网硕互联帮助中心 /* MD / 富文本中的 .toc（含博客园搬家等嵌套结构）；.toc-box 在侧栏，不受影响 */
#content_views .toc,
/* 编辑器常在目录前后插入空 p（:empty 仍占 20px），一并去掉避免顶空隙 */
#content_views.markdown_views > p:empty:has(+ .toc),
#content_views.markdown_views > .toc + p:empty,
/* 富文本旧版目录标记 */
#content_views.htmledit_views #main-toc,
#content_views.htmledit_views #hr-toc,
#content_views.htmledit_views p[id*=\”-toc\”] {
display: none !important;
}
/* 目录去掉后，紧跟的首个标题不再多出一块上边距 */
#content_views.markdown_views > .toc + h1,
#content_views.markdown_views > .toc + h2,
#content_views.markdown_views > .toc + h3,
#content_views.markdown_views > .toc + h4,
#content_views.markdown_views > .toc + p:empty + h1,
#content_views.markdown_views > .toc + p:empty + h2,
#content_views.markdown_views > .toc + p:empty + h3,
#content_views.markdown_views > .toc + p:empty + h4 {
margin-top: 0 !important;
}
智能体发展史：从符号规则到学习型智能体
摘要：现代智能体的形态并不是一次技术飞跃突然决定的，而是在“解决问题、制造新问题、再解决问题”的过程中逐步演化出来的。本文沿着符号主义、专家系统、心智社会、联结主义、强化学习、预训练，最终到 LLM 智能体这条线索，梳理智能体的关键设计思想，并保留一份可直接运行的 ELIZA 代码来观察规则驱动系统的能力边界。
一、先理解演进逻辑：每个范式都在修补上一代的局限
人工智能体几十年的演进有一条很清晰的暗线：每一个新范式的出现，都是为了解决上一代范式的核心痛点；而新方案在带来能力飞跃的同时，又会制造新的局限。
#mermaid-svg-apTZe0opkprSz9B9{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-apTZe0opkprSz9B9 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-apTZe0opkprSz9B9 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-apTZe0opkprSz9B9 .error-icon{fill:#552222;}#mermaid-svg-apTZe0opkprSz9B9 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-apTZe0opkprSz9B9 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-apTZe0opkprSz9B9 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-apTZe0opkprSz9B9 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-apTZe0opkprSz9B9 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-apTZe0opkprSz9B9 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-apTZe0opkprSz9B9 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-apTZe0opkprSz9B9 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-apTZe0opkprSz9B9 .marker.cross{stroke:#333333;}#mermaid-svg-apTZe0opkprSz9B9 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-apTZe0opkprSz9B9 p{margin:0;}#mermaid-svg-apTZe0opkprSz9B9 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-apTZe0opkprSz9B9 .cluster-label text{fill:#333;}#mermaid-svg-apTZe0opkprSz9B9 .cluster-label span{color:#333;}#mermaid-svg-apTZe0opkprSz9B9 .cluster-label span p{background-color:transparent;}#mermaid-svg-apTZe0opkprSz9B9 .label text,#mermaid-svg-apTZe0opkprSz9B9 span{fill:#333;color:#333;}#mermaid-svg-apTZe0opkprSz9B9 .node rect,#mermaid-svg-apTZe0opkprSz9B9 .node circle,#mermaid-svg-apTZe0opkprSz9B9 .node ellipse,#mermaid-svg-apTZe0opkprSz9B9 .node polygon,#mermaid-svg-apTZe0opkprSz9B9 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-apTZe0opkprSz9B9 .rough-node .label text,#mermaid-svg-apTZe0opkprSz9B9 .node .label text,#mermaid-svg-apTZe0opkprSz9B9 .image-shape .label,#mermaid-svg-apTZe0opkprSz9B9 .icon-shape .label{text-anchor:middle;}#mermaid-svg-apTZe0opkprSz9B9 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-apTZe0opkprSz9B9 .rough-node .label,#mermaid-svg-apTZe0opkprSz9B9 .node .label,#mermaid-svg-apTZe0opkprSz9B9 .image-shape .label,#mermaid-svg-apTZe0opkprSz9B9 .icon-shape .label{text-align:center;}#mermaid-svg-apTZe0opkprSz9B9 .node.clickable{cursor:pointer;}#mermaid-svg-apTZe0opkprSz9B9 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-apTZe0opkprSz9B9 .arrowheadPath{fill:#333333;}#mermaid-svg-apTZe0opkprSz9B9 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-apTZe0opkprSz9B9 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-apTZe0opkprSz9B9 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-apTZe0opkprSz9B9 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-apTZe0opkprSz9B9 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-apTZe0opkprSz9B9 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-apTZe0opkprSz9B9 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-apTZe0opkprSz9B9 .cluster text{fill:#333;}#mermaid-svg-apTZe0opkprSz9B9 .cluster span{color:#333;}#mermaid-svg-apTZe0opkprSz9B9 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-apTZe0opkprSz9B9 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-apTZe0opkprSz9B9 rect.text{fill:none;stroke-width:0;}#mermaid-svg-apTZe0opkprSz9B9 .icon-shape,#mermaid-svg-apTZe0opkprSz9B9 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-apTZe0opkprSz9B9 .icon-shape p,#mermaid-svg-apTZe0opkprSz9B9 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-apTZe0opkprSz9B9 .icon-shape .label rect,#mermaid-svg-apTZe0opkprSz9B9 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-apTZe0opkprSz9B9 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-apTZe0opkprSz9B9 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-apTZe0opkprSz9B9 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
符号主义规则 + 逻辑
专家系统知识 + 推理
心智社会去中心化协作
联结主义分布式表示 + 学习
强化学习试错与序贯决策
预训练海量数据 + 世界知识
LLM 智能体感知、规划、工具、行动
把这个链条放在心里，再看每一段技术细节，会更容易理解它们为什么以当时的形态出现，又为什么会被后续方案替代。
二、符号与逻辑：最早期的智能体
早期人工智能研究者相信，人类的智能，尤其是逻辑推理能力，可以被形式化的符号体系捕捉并复现。这种思想后来被称为符号主义，也叫逻辑 AI 或传统 AI。
2.1 物理符号系统假说
符号主义最核心的理论基础，是艾伦·纽厄尔和赫伯特·西蒙提出的物理符号系统假说。它包含两个论断：
- 充分性论断：任何物理符号系统都具备产生通用智能行为的充分手段。
- 必要性论断：任何能够展现通用智能行为的系统，本质上都是一个物理符号系统。
#mermaid-svg-BA5PGXfH4tMFxrRm{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-BA5PGXfH4tMFxrRm .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-BA5PGXfH4tMFxrRm .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-BA5PGXfH4tMFxrRm .error-icon{fill:#552222;}#mermaid-svg-BA5PGXfH4tMFxrRm .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-BA5PGXfH4tMFxrRm .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-BA5PGXfH4tMFxrRm .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-BA5PGXfH4tMFxrRm .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-BA5PGXfH4tMFxrRm .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-BA5PGXfH4tMFxrRm .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-BA5PGXfH4tMFxrRm .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-BA5PGXfH4tMFxrRm .marker{fill:#333333;stroke:#333333;}#mermaid-svg-BA5PGXfH4tMFxrRm .marker.cross{stroke:#333333;}#mermaid-svg-BA5PGXfH4tMFxrRm svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-BA5PGXfH4tMFxrRm p{margin:0;}#mermaid-svg-BA5PGXfH4tMFxrRm .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-BA5PGXfH4tMFxrRm .cluster-label text{fill:#333;}#mermaid-svg-BA5PGXfH4tMFxrRm .cluster-label span{color:#333;}#mermaid-svg-BA5PGXfH4tMFxrRm .cluster-label span p{background-color:transparent;}#mermaid-svg-BA5PGXfH4tMFxrRm .label text,#mermaid-svg-BA5PGXfH4tMFxrRm span{fill:#333;color:#333;}#mermaid-svg-BA5PGXfH4tMFxrRm .node rect,#mermaid-svg-BA5PGXfH4tMFxrRm .node circle,#mermaid-svg-BA5PGXfH4tMFxrRm .node ellipse,#mermaid-svg-BA5PGXfH4tMFxrRm .node polygon,#mermaid-svg-BA5PGXfH4tMFxrRm .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-BA5PGXfH4tMFxrRm .rough-node .label text,#mermaid-svg-BA5PGXfH4tMFxrRm .node .label text,#mermaid-svg-BA5PGXfH4tMFxrRm .image-shape .label,#mermaid-svg-BA5PGXfH4tMFxrRm .icon-shape .label{text-anchor:middle;}#mermaid-svg-BA5PGXfH4tMFxrRm .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-BA5PGXfH4tMFxrRm .rough-node .label,#mermaid-svg-BA5PGXfH4tMFxrRm .node .label,#mermaid-svg-BA5PGXfH4tMFxrRm .image-shape .label,#mermaid-svg-BA5PGXfH4tMFxrRm .icon-shape .label{text-align:center;}#mermaid-svg-BA5PGXfH4tMFxrRm .node.clickable{cursor:pointer;}#mermaid-svg-BA5PGXfH4tMFxrRm .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-BA5PGXfH4tMFxrRm .arrowheadPath{fill:#333333;}#mermaid-svg-BA5PGXfH4tMFxrRm .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-BA5PGXfH4tMFxrRm .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-BA5PGXfH4tMFxrRm .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-BA5PGXfH4tMFxrRm .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-BA5PGXfH4tMFxrRm .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-BA5PGXfH4tMFxrRm .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-BA5PGXfH4tMFxrRm .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-BA5PGXfH4tMFxrRm .cluster text{fill:#333;}#mermaid-svg-BA5PGXfH4tMFxrRm .cluster span{color:#333;}#mermaid-svg-BA5PGXfH4tMFxrRm div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-BA5PGXfH4tMFxrRm .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-BA5PGXfH4tMFxrRm rect.text{fill:none;stroke-width:0;}#mermaid-svg-BA5PGXfH4tMFxrRm .icon-shape,#mermaid-svg-BA5PGXfH4tMFxrRm .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-BA5PGXfH4tMFxrRm .icon-shape p,#mermaid-svg-BA5PGXfH4tMFxrRm .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-BA5PGXfH4tMFxrRm .icon-shape .label rect,#mermaid-svg-BA5PGXfH4tMFxrRm .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-BA5PGXfH4tMFxrRm .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-BA5PGXfH4tMFxrRm .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-BA5PGXfH4tMFxrRm :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
物理符号系统假说
充分性：符号系统足以产生通用智能
必要性：通用智能本质上是符号系统
工程目标：知识表示 + 推理算法
这套假说把“心智是什么”这样一个模糊的哲学问题，转化成了一个可以用知识表示和推理算法来处理的工程问题。专家系统和自动规划，几乎都是在这个框架下展开的。
2.2 专家系统：知识库与推理机分离
专家系统是符号主义最成功的应用形态。它的目标是模拟领域专家解决问题的能力：把专家的知识和经验编码进程序，在遇到相似问题时给出接近专家水平的结论或建议。
#mermaid-svg-Utsf41H2UUOmsDtX{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-Utsf41H2UUOmsDtX .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-Utsf41H2UUOmsDtX .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-Utsf41H2UUOmsDtX .error-icon{fill:#552222;}#mermaid-svg-Utsf41H2UUOmsDtX .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-Utsf41H2UUOmsDtX .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-Utsf41H2UUOmsDtX .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-Utsf41H2UUOmsDtX .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-Utsf41H2UUOmsDtX .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-Utsf41H2UUOmsDtX .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-Utsf41H2UUOmsDtX .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-Utsf41H2UUOmsDtX .marker{fill:#333333;stroke:#333333;}#mermaid-svg-Utsf41H2UUOmsDtX .marker.cross{stroke:#333333;}#mermaid-svg-Utsf41H2UUOmsDtX svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-Utsf41H2UUOmsDtX p{margin:0;}#mermaid-svg-Utsf41H2UUOmsDtX .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-Utsf41H2UUOmsDtX .cluster-label text{fill:#333;}#mermaid-svg-Utsf41H2UUOmsDtX .cluster-label span{color:#333;}#mermaid-svg-Utsf41H2UUOmsDtX .cluster-label span p{background-color:transparent;}#mermaid-svg-Utsf41H2UUOmsDtX .label text,#mermaid-svg-Utsf41H2UUOmsDtX span{fill:#333;color:#333;}#mermaid-svg-Utsf41H2UUOmsDtX .node rect,#mermaid-svg-Utsf41H2UUOmsDtX .node circle,#mermaid-svg-Utsf41H2UUOmsDtX .node ellipse,#mermaid-svg-Utsf41H2UUOmsDtX .node polygon,#mermaid-svg-Utsf41H2UUOmsDtX .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-Utsf41H2UUOmsDtX .rough-node .label text,#mermaid-svg-Utsf41H2UUOmsDtX .node .label text,#mermaid-svg-Utsf41H2UUOmsDtX .image-shape .label,#mermaid-svg-Utsf41H2UUOmsDtX .icon-shape .label{text-anchor:middle;}#mermaid-svg-Utsf41H2UUOmsDtX .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-Utsf41H2UUOmsDtX .rough-node .label,#mermaid-svg-Utsf41H2UUOmsDtX .node .label,#mermaid-svg-Utsf41H2UUOmsDtX .image-shape .label,#mermaid-svg-Utsf41H2UUOmsDtX .icon-shape .label{text-align:center;}#mermaid-svg-Utsf41H2UUOmsDtX .node.clickable{cursor:pointer;}#mermaid-svg-Utsf41H2UUOmsDtX .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-Utsf41H2UUOmsDtX .arrowheadPath{fill:#333333;}#mermaid-svg-Utsf41H2UUOmsDtX .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-Utsf41H2UUOmsDtX .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-Utsf41H2UUOmsDtX .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-Utsf41H2UUOmsDtX .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-Utsf41H2UUOmsDtX .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-Utsf41H2UUOmsDtX .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-Utsf41H2UUOmsDtX .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-Utsf41H2UUOmsDtX .cluster text{fill:#333;}#mermaid-svg-Utsf41H2UUOmsDtX .cluster span{color:#333;}#mermaid-svg-Utsf41H2UUOmsDtX div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-Utsf41H2UUOmsDtX .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-Utsf41H2UUOmsDtX rect.text{fill:none;stroke-width:0;}#mermaid-svg-Utsf41H2UUOmsDtX .icon-shape,#mermaid-svg-Utsf41H2UUOmsDtX .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-Utsf41H2UUOmsDtX .icon-shape p,#mermaid-svg-Utsf41H2UUOmsDtX .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-Utsf41H2UUOmsDtX .icon-shape .label rect,#mermaid-svg-Utsf41H2UUOmsDtX .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-Utsf41H2UUOmsDtX .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-Utsf41H2UUOmsDtX .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-Utsf41H2UUOmsDtX :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
用户
用户界面
推理机
知识库IF-THEN 规则
工作记忆 / 当前事实
专家系统的关键设计，是把“知识”和“推理”分离。
- 知识库负责存放领域知识，最常见的形式是产生式规则，也就是一系列 IF-THEN 条件语句。
- 推理机负责根据当前事实寻找并应用规则，从而推导新结论。
推理机主要有两种工作方式：
#mermaid-svg-vzY7hwzz7uf6XwPs{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-vzY7hwzz7uf6XwPs .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-vzY7hwzz7uf6XwPs .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-vzY7hwzz7uf6XwPs .error-icon{fill:#552222;}#mermaid-svg-vzY7hwzz7uf6XwPs .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-vzY7hwzz7uf6XwPs .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-vzY7hwzz7uf6XwPs .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-vzY7hwzz7uf6XwPs .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-vzY7hwzz7uf6XwPs .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-vzY7hwzz7uf6XwPs .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-vzY7hwzz7uf6XwPs .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-vzY7hwzz7uf6XwPs .marker{fill:#333333;stroke:#333333;}#mermaid-svg-vzY7hwzz7uf6XwPs .marker.cross{stroke:#333333;}#mermaid-svg-vzY7hwzz7uf6XwPs svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-vzY7hwzz7uf6XwPs p{margin:0;}#mermaid-svg-vzY7hwzz7uf6XwPs .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-vzY7hwzz7uf6XwPs .cluster-label text{fill:#333;}#mermaid-svg-vzY7hwzz7uf6XwPs .cluster-label span{color:#333;}#mermaid-svg-vzY7hwzz7uf6XwPs .cluster-label span p{background-color:transparent;}#mermaid-svg-vzY7hwzz7uf6XwPs .label text,#mermaid-svg-vzY7hwzz7uf6XwPs span{fill:#333;color:#333;}#mermaid-svg-vzY7hwzz7uf6XwPs .node rect,#mermaid-svg-vzY7hwzz7uf6XwPs .node circle,#mermaid-svg-vzY7hwzz7uf6XwPs .node ellipse,#mermaid-svg-vzY7hwzz7uf6XwPs .node polygon,#mermaid-svg-vzY7hwzz7uf6XwPs .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-vzY7hwzz7uf6XwPs .rough-node .label text,#mermaid-svg-vzY7hwzz7uf6XwPs .node .label text,#mermaid-svg-vzY7hwzz7uf6XwPs .image-shape .label,#mermaid-svg-vzY7hwzz7uf6XwPs .icon-shape .label{text-anchor:middle;}#mermaid-svg-vzY7hwzz7uf6XwPs .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-vzY7hwzz7uf6XwPs .rough-node .label,#mermaid-svg-vzY7hwzz7uf6XwPs .node .label,#mermaid-svg-vzY7hwzz7uf6XwPs .image-shape .label,#mermaid-svg-vzY7hwzz7uf6XwPs .icon-shape .label{text-align:center;}#mermaid-svg-vzY7hwzz7uf6XwPs .node.clickable{cursor:pointer;}#mermaid-svg-vzY7hwzz7uf6XwPs .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-vzY7hwzz7uf6XwPs .arrowheadPath{fill:#333333;}#mermaid-svg-vzY7hwzz7uf6XwPs .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-vzY7hwzz7uf6XwPs .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-vzY7hwzz7uf6XwPs .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-vzY7hwzz7uf6XwPs .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-vzY7hwzz7uf6XwPs .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-vzY7hwzz7uf6XwPs .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-vzY7hwzz7uf6XwPs .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-vzY7hwzz7uf6XwPs .cluster text{fill:#333;}#mermaid-svg-vzY7hwzz7uf6XwPs .cluster span{color:#333;}#mermaid-svg-vzY7hwzz7uf6XwPs div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-vzY7hwzz7uf6XwPs .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-vzY7hwzz7uf6XwPs rect.text{fill:none;stroke-width:0;}#mermaid-svg-vzY7hwzz7uf6XwPs .icon-shape,#mermaid-svg-vzY7hwzz7uf6XwPs .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-vzY7hwzz7uf6XwPs .icon-shape p,#mermaid-svg-vzY7hwzz7uf6XwPs .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-vzY7hwzz7uf6XwPs .icon-shape .label rect,#mermaid-svg-vzY7hwzz7uf6XwPs .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-vzY7hwzz7uf6XwPs .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-vzY7hwzz7uf6XwPs .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-vzY7hwzz7uf6XwPs :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
正向链：匹配 IF
触发 THEN
反向链：寻找能证明目标的规则
已知事实
产生式规则
新结论
假设目标
- 正向链从事实出发，数据驱动地推出新结论。
- 反向链从假设目标出发，目标驱动地寻找支持证据。
最典型的案例是 MYCIN 系统。它通过问答交互收集症状、病史和化验结果，用约 600 条规则辅助诊断细菌性血液感染。MYCIN 还引入了置信因子，用 -1 到 1 的数值表示结论可信度，使系统能够处理医学诊断中的不确定性。
#mermaid-svg-vltzi0shpynx9Paa{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-vltzi0shpynx9Paa .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-vltzi0shpynx9Paa .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-vltzi0shpynx9Paa .error-icon{fill:#552222;}#mermaid-svg-vltzi0shpynx9Paa .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-vltzi0shpynx9Paa .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-vltzi0shpynx9Paa .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-vltzi0shpynx9Paa .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-vltzi0shpynx9Paa .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-vltzi0shpynx9Paa .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-vltzi0shpynx9Paa .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-vltzi0shpynx9Paa .marker{fill:#333333;stroke:#333333;}#mermaid-svg-vltzi0shpynx9Paa .marker.cross{stroke:#333333;}#mermaid-svg-vltzi0shpynx9Paa svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-vltzi0shpynx9Paa p{margin:0;}#mermaid-svg-vltzi0shpynx9Paa .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-vltzi0shpynx9Paa .cluster-label text{fill:#333;}#mermaid-svg-vltzi0shpynx9Paa .cluster-label span{color:#333;}#mermaid-svg-vltzi0shpynx9Paa .cluster-label span p{background-color:transparent;}#mermaid-svg-vltzi0shpynx9Paa .label text,#mermaid-svg-vltzi0shpynx9Paa span{fill:#333;color:#333;}#mermaid-svg-vltzi0shpynx9Paa .node rect,#mermaid-svg-vltzi0shpynx9Paa .node circle,#mermaid-svg-vltzi0shpynx9Paa .node ellipse,#mermaid-svg-vltzi0shpynx9Paa .node polygon,#mermaid-svg-vltzi0shpynx9Paa .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-vltzi0shpynx9Paa .rough-node .label text,#mermaid-svg-vltzi0shpynx9Paa .node .label text,#mermaid-svg-vltzi0shpynx9Paa .image-shape .label,#mermaid-svg-vltzi0shpynx9Paa .icon-shape .label{text-anchor:middle;}#mermaid-svg-vltzi0shpynx9Paa .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-vltzi0shpynx9Paa .rough-node .label,#mermaid-svg-vltzi0shpynx9Paa .node .label,#mermaid-svg-vltzi0shpynx9Paa .image-shape .label,#mermaid-svg-vltzi0shpynx9Paa .icon-shape .label{text-align:center;}#mermaid-svg-vltzi0shpynx9Paa .node.clickable{cursor:pointer;}#mermaid-svg-vltzi0shpynx9Paa .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-vltzi0shpynx9Paa .arrowheadPath{fill:#333333;}#mermaid-svg-vltzi0shpynx9Paa .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-vltzi0shpynx9Paa .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-vltzi0shpynx9Paa .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-vltzi0shpynx9Paa .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-vltzi0shpynx9Paa .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-vltzi0shpynx9Paa .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-vltzi0shpynx9Paa .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-vltzi0shpynx9Paa .cluster text{fill:#333;}#mermaid-svg-vltzi0shpynx9Paa .cluster span{color:#333;}#mermaid-svg-vltzi0shpynx9Paa div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-vltzi0shpynx9Paa .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-vltzi0shpynx9Paa rect.text{fill:none;stroke-width:0;}#mermaid-svg-vltzi0shpynx9Paa .icon-shape,#mermaid-svg-vltzi0shpynx9Paa .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-vltzi0shpynx9Paa .icon-shape p,#mermaid-svg-vltzi0shpynx9Paa .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-vltzi0shpynx9Paa .icon-shape .label rect,#mermaid-svg-vltzi0shpynx9Paa .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-vltzi0shpynx9Paa .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-vltzi0shpynx9Paa .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-vltzi0shpynx9Paa :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
最高目标：确定致病菌
反向寻找所需证据
向医生提问获取信息
规则匹配与不确定性计算
输出诊断与置信度
2.3 SHRDLU：一次小型综合智能实验
SHRDLU 是特里·威诺格拉德在 1968 到 1970 年开发的程序。它在一个由积木和虚拟机械臂组成的封闭世界里，尝试把自然语言理解、规划、行动和记忆放进同一个系统。
#mermaid-svg-mkNHsbiuHpp8U7gK{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-mkNHsbiuHpp8U7gK .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-mkNHsbiuHpp8U7gK .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-mkNHsbiuHpp8U7gK .error-icon{fill:#552222;}#mermaid-svg-mkNHsbiuHpp8U7gK .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-mkNHsbiuHpp8U7gK .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-mkNHsbiuHpp8U7gK .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-mkNHsbiuHpp8U7gK .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-mkNHsbiuHpp8U7gK .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-mkNHsbiuHpp8U7gK .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-mkNHsbiuHpp8U7gK .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-mkNHsbiuHpp8U7gK .marker{fill:#333333;stroke:#333333;}#mermaid-svg-mkNHsbiuHpp8U7gK .marker.cross{stroke:#333333;}#mermaid-svg-mkNHsbiuHpp8U7gK svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-mkNHsbiuHpp8U7gK p{margin:0;}#mermaid-svg-mkNHsbiuHpp8U7gK .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-mkNHsbiuHpp8U7gK .cluster-label text{fill:#333;}#mermaid-svg-mkNHsbiuHpp8U7gK .cluster-label span{color:#333;}#mermaid-svg-mkNHsbiuHpp8U7gK .cluster-label span p{background-color:transparent;}#mermaid-svg-mkNHsbiuHpp8U7gK .label text,#mermaid-svg-mkNHsbiuHpp8U7gK span{fill:#333;color:#333;}#mermaid-svg-mkNHsbiuHpp8U7gK .node rect,#mermaid-svg-mkNHsbiuHpp8U7gK .node circle,#mermaid-svg-mkNHsbiuHpp8U7gK .node ellipse,#mermaid-svg-mkNHsbiuHpp8U7gK .node polygon,#mermaid-svg-mkNHsbiuHpp8U7gK .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-mkNHsbiuHpp8U7gK .rough-node .label text,#mermaid-svg-mkNHsbiuHpp8U7gK .node .label text,#mermaid-svg-mkNHsbiuHpp8U7gK .image-shape .label,#mermaid-svg-mkNHsbiuHpp8U7gK .icon-shape .label{text-anchor:middle;}#mermaid-svg-mkNHsbiuHpp8U7gK .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-mkNHsbiuHpp8U7gK .rough-node .label,#mermaid-svg-mkNHsbiuHpp8U7gK .node .label,#mermaid-svg-mkNHsbiuHpp8U7gK .image-shape .label,#mermaid-svg-mkNHsbiuHpp8U7gK .icon-shape .label{text-align:center;}#mermaid-svg-mkNHsbiuHpp8U7gK .node.clickable{cursor:pointer;}#mermaid-svg-mkNHsbiuHpp8U7gK .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-mkNHsbiuHpp8U7gK .arrowheadPath{fill:#333333;}#mermaid-svg-mkNHsbiuHpp8U7gK .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-mkNHsbiuHpp8U7gK .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-mkNHsbiuHpp8U7gK .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-mkNHsbiuHpp8U7gK .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-mkNHsbiuHpp8U7gK .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-mkNHsbiuHpp8U7gK .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-mkNHsbiuHpp8U7gK .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-mkNHsbiuHpp8U7gK .cluster text{fill:#333;}#mermaid-svg-mkNHsbiuHpp8U7gK .cluster span{color:#333;}#mermaid-svg-mkNHsbiuHpp8U7gK div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-mkNHsbiuHpp8U7gK .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-mkNHsbiuHpp8U7gK rect.text{fill:none;stroke-width:0;}#mermaid-svg-mkNHsbiuHpp8U7gK .icon-shape,#mermaid-svg-mkNHsbiuHpp8U7gK .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-mkNHsbiuHpp8U7gK .icon-shape p,#mermaid-svg-mkNHsbiuHpp8U7gK .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-mkNHsbiuHpp8U7gK .icon-shape .label rect,#mermaid-svg-mkNHsbiuHpp8U7gK .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-mkNHsbiuHpp8U7gK .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-mkNHsbiuHpp8U7gK .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-mkNHsbiuHpp8U7gK :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
自然语言输入
语言理解
规划行动序列
机械臂执行动作
积木世界状态
环境与行为记忆
它能够处理指代消解和上下文记忆。例如：
- 指代消解：Find a block which is taller than the one you are holding and put it into the box.
- 上下文记忆：先说 Grasp the pyramid.，再问 What does the box contain?
- 行为历史：问它 Why did you pick up the red block?，它可以回答 BECAUSE YOU ASKED ME TO.
SHRDLU 的重要性不在于它走出了积木世界，而在于它第一次把多个独立 AI 模块组织成了一个统一的“感知、思考、行动”闭环。这个结构成为后来智能体研究的重要原型。
2.4 符号主义为什么难以走向开放世界
从 20 世纪 80 年代起，符号主义 AI 从封闭的微观世界走向现实世界时，遇到两类根本性困难。
知识获取与常识问题
专家系统的能力完全依赖知识库。但人工整理专家知识成本极高、难以规模化，而且很多专家知识是内隐和直觉性的，很难写成 IF-THEN 规则。即使建立了庞大知识库，系统对“水是湿的”“绳子可以拉但不能推”这类常识也一无所知，除非规则被显式写进去。
框架问题与系统脆弱性
在动态世界中，智能体执行一个动作后，如何判断哪些事物没有变化，这是一个逻辑难题。更棘手的是脆弱性：符号系统只认识预设规则，遇到规则之外的小变化就可能完全失灵。
SHRDLU 之所以成功，恰恰是因为它运行在一个规则完备的封闭世界。真实世界则充满例外。
三、动手观察：用规则实现一个 ELIZA 聊天机器人
为了更直观地理解规则驱动系统的能力和局限，我们可以复现历史上很有影响力的早期聊天机器人 ELIZA。
3.1 ELIZA 的设计思想
ELIZA 由约瑟夫·魏泽鲍姆于 1966 年发布。它并不真正理解内容，而是识别关键词，再用预设转换规则把用户的话改写成开放式提问。最有名的脚本 DOCTOR 模仿了一位非指导性心理治疗师。
例如，用户说“我为我的男朋友感到难过”，ELIZA 可能识别到“我为……感到难过”，并回应：“你为什么会为你的男朋友感到难过？”
这个设计想证明的并不是机器能够理解人类，而是简单的句式转换就可以制造一种“智能”和“共情”的假象。
3.2 模式匹配与文本替换
ELIZA 的工作流程可以概括为关键词识别、分解、重组、代词转换四步。
下面保留原文档中的伪代码思路：
FUNCTION generate_response(user_input):
// 1. 将用户输入拆分成单词
words = SPLIT(user_input)
// 2. 寻找优先级最高的关键词规则
best_rule = FIND_BEST_RULE(words)
IF best_rule is NULL:
RETURN a_generic_response() // 例如:"Please go on."
// 3. 使用规则分解用户输入
decomposed_parts = DECOMPOSE(user_input, best_rule.decomposition_pattern)
IF decomposition_failed:
RETURN a_generic_response()
// 4. 对分解出的部分进行代词转换
transformed_parts = TRANSFORM_PRONOUNS(decomposed_parts)
// 5. 使用重组规则生成回应
response = REASSEMBLE(transformed_parts, best_rule.reassembly_patterns)
RETURN response
通过这套机制，复杂的自然语言理解问题被简化成一个基于规则的模式匹配游戏。
3.3 核心逻辑实现
下面是迷你版 ELIZA，它只包含一小部分规则，但足以展示核心工作机制。
import re
import random
# 定义规则库:模式(正则表达式) -> 响应模板列表
rules = {
r'I need (.*)': [
"Why do you need {0}?",
"Would it really help you to get {0}?",
"Are you sure you need {0}?"
],
r'Why don\\'t you (.*)\\?': [
"Do you really think I don't {0}?",
"Perhaps eventually I will {0}.",
"Do you really want me to {0}?"
],
r'Why can\\'t I (.*)\\?': [
"Do you think you should be able to {0}?",
"If you could {0}, what would you do?",
"I don't know — why can't you {0}?"
],
r'I am (.*)': [
"Did you come to me because you are {0}?",
"How long have you been {0}?",
"How do you feel about being {0}?"
],
r'.* mother .*': [
"Tell me more about your mother.",
"What was your relationship with your mother like?",
"How do you feel about your mother?"
],
r'.* father .*': [
"Tell me more about your father.",
"How did your father make you feel?",
"What has your father taught you?"
],
r'.*': [
"Please tell me more.",
"Let's change focus a bit… Tell me about your family.",
"Can you elaborate on that?"
]
}
# 定义代词转换规则
pronoun_swap = {
"i": "you", "you": "i", "me": "you", "my": "your",
"am": "are", "are": "am", "was": "were", "i'd": "you would",
"i've": "you have", "i'll": "you will", "yours": "mine",
"mine": "yours"
}
def swap_pronouns(phrase):
"""
对输入短语中的代词进行第一/第二人称转换
"""
words = phrase.lower().split()
swapped_words = [pronoun_swap.get(word, word) for word in words]
return " ".join(swapped_words)
def respond(user_input):
"""
根据规则库生成响应
"""
for pattern, responses in rules.items():
match = re.search(pattern, user_input, re.IGNORECASE)
if match:
# 捕获匹配到的部分
captured_group = match.group(1) if match.groups() else ''
# 进行代词转换
swapped_group = swap_pronouns(captured_group)
# 从模板中随机选择一个并格式化
response = random.choice(responses).format(swapped_group)
return response
# 如果没有匹配任何特定规则，使用最后的通配符规则
return random.choice(rules[r'.*'])
# 主聊天循环
if __name__ == '__main__':
print("Therapist: Hello! How can I help you today?")
while True:
user_input = input("You: ")
if user_input.lower() in ["quit", "exit", "bye"]:
print("Therapist: Goodbye. It was nice talking to you.")
break
response = respond(user_input)
print(f"Therapist: {response}")
>>>
Therapist: Hello! How can I help you today?
You: I am feeling sad today.
Therapist: How long have you been feeling sad today?
You: I need some help with my project.
Therapist: Are you sure you need some help with your project?
You: My mother is not happy with my work.
Therapist: Tell me more about your mother.
You: quit
Therapist: Goodbye. It was nice talking to you.
3.4 从 ELIZA 看到规则驱动系统的边界
这个实现把符号主义的优势与局限都暴露得很清楚。
- 缺乏语义理解：面对 I am not happy，系统仍可能机械匹配 I am (.*) 并给出语义不通的回应，因为它不理解否定词的作用。
- 无上下文记忆：每次回应只基于当前单句输入，系统是无状态的。
- 规则扩展困难：规则数量增加后，冲突、优先级和维护成本会迅速上升。
不过，当时许多用户仍然相信 ELIZA 能理解自己，这就是 ELIZA 效应。它说明“看似智能”并不等于“真正理解”。
#mermaid-svg-Low4FkAswHcePKzE{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-Low4FkAswHcePKzE .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-Low4FkAswHcePKzE .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-Low4FkAswHcePKzE .error-icon{fill:#552222;}#mermaid-svg-Low4FkAswHcePKzE .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-Low4FkAswHcePKzE .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-Low4FkAswHcePKzE .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-Low4FkAswHcePKzE .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-Low4FkAswHcePKzE .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-Low4FkAswHcePKzE .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-Low4FkAswHcePKzE .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-Low4FkAswHcePKzE .marker{fill:#333333;stroke:#333333;}#mermaid-svg-Low4FkAswHcePKzE .marker.cross{stroke:#333333;}#mermaid-svg-Low4FkAswHcePKzE svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-Low4FkAswHcePKzE p{margin:0;}#mermaid-svg-Low4FkAswHcePKzE .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-Low4FkAswHcePKzE .cluster-label text{fill:#333;}#mermaid-svg-Low4FkAswHcePKzE .cluster-label span{color:#333;}#mermaid-svg-Low4FkAswHcePKzE .cluster-label span p{background-color:transparent;}#mermaid-svg-Low4FkAswHcePKzE .label text,#mermaid-svg-Low4FkAswHcePKzE span{fill:#333;color:#333;}#mermaid-svg-Low4FkAswHcePKzE .node rect,#mermaid-svg-Low4FkAswHcePKzE .node circle,#mermaid-svg-Low4FkAswHcePKzE .node ellipse,#mermaid-svg-Low4FkAswHcePKzE .node polygon,#mermaid-svg-Low4FkAswHcePKzE .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-Low4FkAswHcePKzE .rough-node .label text,#mermaid-svg-Low4FkAswHcePKzE .node .label text,#mermaid-svg-Low4FkAswHcePKzE .image-shape .label,#mermaid-svg-Low4FkAswHcePKzE .icon-shape .label{text-anchor:middle;}#mermaid-svg-Low4FkAswHcePKzE .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-Low4FkAswHcePKzE .rough-node .label,#mermaid-svg-Low4FkAswHcePKzE .node .label,#mermaid-svg-Low4FkAswHcePKzE .image-shape .label,#mermaid-svg-Low4FkAswHcePKzE .icon-shape .label{text-align:center;}#mermaid-svg-Low4FkAswHcePKzE .node.clickable{cursor:pointer;}#mermaid-svg-Low4FkAswHcePKzE .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-Low4FkAswHcePKzE .arrowheadPath{fill:#333333;}#mermaid-svg-Low4FkAswHcePKzE .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-Low4FkAswHcePKzE .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-Low4FkAswHcePKzE .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-Low4FkAswHcePKzE .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-Low4FkAswHcePKzE .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-Low4FkAswHcePKzE .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-Low4FkAswHcePKzE .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-Low4FkAswHcePKzE .cluster text{fill:#333;}#mermaid-svg-Low4FkAswHcePKzE .cluster span{color:#333;}#mermaid-svg-Low4FkAswHcePKzE div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-Low4FkAswHcePKzE .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-Low4FkAswHcePKzE rect.text{fill:none;stroke-width:0;}#mermaid-svg-Low4FkAswHcePKzE .icon-shape,#mermaid-svg-Low4FkAswHcePKzE .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-Low4FkAswHcePKzE .icon-shape p,#mermaid-svg-Low4FkAswHcePKzE .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-Low4FkAswHcePKzE .icon-shape .label rect,#mermaid-svg-Low4FkAswHcePKzE .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-Low4FkAswHcePKzE .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-Low4FkAswHcePKzE .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-Low4FkAswHcePKzE :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
用户输入
关键词与模式匹配
代词转换
重组模板
看似连贯的回应
四、心智社会：从单一大脑到去中心化协作
符号主义通过预设规则构建了一个集中式推理引擎，但面对真实世界的模糊和无穷变化时，它始终显得脆弱。马文·明斯基在《心智社会》中提出了另一种思路：智能不是来自某个完美原则，而是来自大量简单、专门化心智过程之间的协作。
4.1 把心智看作一个社会
在明斯基的框架里，智能体不是今天常见的 LLM 智能体，而是一个极其简单、专门化、自身“无心”的过程。例如 LINE-FINDER 负责识别线条，GRASP 负责抓握。
这些简单智能体组成机构，机构再通过去中心化的激活和抑制信号相互影响。复杂行为不是由某个高级智能体预先规划的，而是从局部交互中涌现出来的。
#mermaid-svg-NOugM70Pnr44uRN5{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-NOugM70Pnr44uRN5 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-NOugM70Pnr44uRN5 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-NOugM70Pnr44uRN5 .error-icon{fill:#552222;}#mermaid-svg-NOugM70Pnr44uRN5 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-NOugM70Pnr44uRN5 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-NOugM70Pnr44uRN5 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-NOugM70Pnr44uRN5 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-NOugM70Pnr44uRN5 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-NOugM70Pnr44uRN5 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-NOugM70Pnr44uRN5 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-NOugM70Pnr44uRN5 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-NOugM70Pnr44uRN5 .marker.cross{stroke:#333333;}#mermaid-svg-NOugM70Pnr44uRN5 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-NOugM70Pnr44uRN5 p{margin:0;}#mermaid-svg-NOugM70Pnr44uRN5 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-NOugM70Pnr44uRN5 .cluster-label text{fill:#333;}#mermaid-svg-NOugM70Pnr44uRN5 .cluster-label span{color:#333;}#mermaid-svg-NOugM70Pnr44uRN5 .cluster-label span p{background-color:transparent;}#mermaid-svg-NOugM70Pnr44uRN5 .label text,#mermaid-svg-NOugM70Pnr44uRN5 span{fill:#333;color:#333;}#mermaid-svg-NOugM70Pnr44uRN5 .node rect,#mermaid-svg-NOugM70Pnr44uRN5 .node circle,#mermaid-svg-NOugM70Pnr44uRN5 .node ellipse,#mermaid-svg-NOugM70Pnr44uRN5 .node polygon,#mermaid-svg-NOugM70Pnr44uRN5 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-NOugM70Pnr44uRN5 .rough-node .label text,#mermaid-svg-NOugM70Pnr44uRN5 .node .label text,#mermaid-svg-NOugM70Pnr44uRN5 .image-shape .label,#mermaid-svg-NOugM70Pnr44uRN5 .icon-shape .label{text-anchor:middle;}#mermaid-svg-NOugM70Pnr44uRN5 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-NOugM70Pnr44uRN5 .rough-node .label,#mermaid-svg-NOugM70Pnr44uRN5 .node .label,#mermaid-svg-NOugM70Pnr44uRN5 .image-shape .label,#mermaid-svg-NOugM70Pnr44uRN5 .icon-shape .label{text-align:center;}#mermaid-svg-NOugM70Pnr44uRN5 .node.clickable{cursor:pointer;}#mermaid-svg-NOugM70Pnr44uRN5 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-NOugM70Pnr44uRN5 .arrowheadPath{fill:#333333;}#mermaid-svg-NOugM70Pnr44uRN5 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-NOugM70Pnr44uRN5 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-NOugM70Pnr44uRN5 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-NOugM70Pnr44uRN5 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-NOugM70Pnr44uRN5 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-NOugM70Pnr44uRN5 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-NOugM70Pnr44uRN5 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-NOugM70Pnr44uRN5 .cluster text{fill:#333;}#mermaid-svg-NOugM70Pnr44uRN5 .cluster span{color:#333;}#mermaid-svg-NOugM70Pnr44uRN5 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-NOugM70Pnr44uRN5 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-NOugM70Pnr44uRN5 rect.text{fill:none;stroke-width:0;}#mermaid-svg-NOugM70Pnr44uRN5 .icon-shape,#mermaid-svg-NOugM70Pnr44uRN5 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-NOugM70Pnr44uRN5 .icon-shape p,#mermaid-svg-NOugM70Pnr44uRN5 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-NOugM70Pnr44uRN5 .icon-shape .label rect,#mermaid-svg-NOugM70Pnr44uRN5 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-NOugM70Pnr44uRN5 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-NOugM70Pnr44uRN5 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-NOugM70Pnr44uRN5 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
BUILD-TOWER
BUILDER
ADD-BLOCK
FIND-BLOCK
GET-BLOCK
PUT-ON-TOP
SEE-SHAPE
REACH
GRASP
以“搭建积木塔”为例：BUILD-TOWER 只负责激活下属机构，BUILDER 只负责循环，ADD-BLOCK 协调寻找、抓取和放置积木。没有任何一个组件掌握完整计划，但协作结果看起来像一个完整智能行为。
4.2 对多智能体系统的启发
心智社会理论为分布式人工智能和多智能体系统提供了重要概念基础。
- 去中心化控制：不存在中央控制器，协调机制和任务分配需要依靠局部交互完成。
- 涌现式计算：复杂问题的解可以从简单局部规则中自发产生，例如蚁群算法和粒子群优化。
- 智能体的社会性：智能体之间的交互被扩展为通信语言、交互协议、协商策略和组织结构。
这标志着研究焦点从“如何构建一个全能的单一智能体”转向“如何设计一个高效协作的智能体群体”。
五、学习范式的演进与现代智能体
如果智能无法被完全设计出来，那它是否可以学习出来？这个问题开启了人工智能的学习时代。
5.1 从符号到联结：知识不再被手工写入
联结主义与符号主义走了相反的方向。它不把知识写进规则库，而是用连接权重分布式地保存在大量简单人工神经元之间。系统的智能来自学习过程，通过接触样本和反向传播不断调整权重。
#mermaid-svg-ClnWffHW25PT4kII{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-ClnWffHW25PT4kII .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-ClnWffHW25PT4kII .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-ClnWffHW25PT4kII .error-icon{fill:#552222;}#mermaid-svg-ClnWffHW25PT4kII .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-ClnWffHW25PT4kII .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-ClnWffHW25PT4kII .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-ClnWffHW25PT4kII .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-ClnWffHW25PT4kII .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-ClnWffHW25PT4kII .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-ClnWffHW25PT4kII .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-ClnWffHW25PT4kII .marker{fill:#333333;stroke:#333333;}#mermaid-svg-ClnWffHW25PT4kII .marker.cross{stroke:#333333;}#mermaid-svg-ClnWffHW25PT4kII svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-ClnWffHW25PT4kII p{margin:0;}#mermaid-svg-ClnWffHW25PT4kII .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-ClnWffHW25PT4kII .cluster-label text{fill:#333;}#mermaid-svg-ClnWffHW25PT4kII .cluster-label span{color:#333;}#mermaid-svg-ClnWffHW25PT4kII .cluster-label span p{background-color:transparent;}#mermaid-svg-ClnWffHW25PT4kII .label text,#mermaid-svg-ClnWffHW25PT4kII span{fill:#333;color:#333;}#mermaid-svg-ClnWffHW25PT4kII .node rect,#mermaid-svg-ClnWffHW25PT4kII .node circle,#mermaid-svg-ClnWffHW25PT4kII .node ellipse,#mermaid-svg-ClnWffHW25PT4kII .node polygon,#mermaid-svg-ClnWffHW25PT4kII .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-ClnWffHW25PT4kII .rough-node .label text,#mermaid-svg-ClnWffHW25PT4kII .node .label text,#mermaid-svg-ClnWffHW25PT4kII .image-shape .label,#mermaid-svg-ClnWffHW25PT4kII .icon-shape .label{text-anchor:middle;}#mermaid-svg-ClnWffHW25PT4kII .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-ClnWffHW25PT4kII .rough-node .label,#mermaid-svg-ClnWffHW25PT4kII .node .label,#mermaid-svg-ClnWffHW25PT4kII .image-shape .label,#mermaid-svg-ClnWffHW25PT4kII .icon-shape .label{text-align:center;}#mermaid-svg-ClnWffHW25PT4kII .node.clickable{cursor:pointer;}#mermaid-svg-ClnWffHW25PT4kII .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-ClnWffHW25PT4kII .arrowheadPath{fill:#333333;}#mermaid-svg-ClnWffHW25PT4kII .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-ClnWffHW25PT4kII .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-ClnWffHW25PT4kII .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-ClnWffHW25PT4kII .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-ClnWffHW25PT4kII .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-ClnWffHW25PT4kII .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-ClnWffHW25PT4kII .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-ClnWffHW25PT4kII .cluster text{fill:#333;}#mermaid-svg-ClnWffHW25PT4kII .cluster span{color:#333;}#mermaid-svg-ClnWffHW25PT4kII div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-ClnWffHW25PT4kII .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-ClnWffHW25PT4kII rect.text{fill:none;stroke-width:0;}#mermaid-svg-ClnWffHW25PT4kII .icon-shape,#mermaid-svg-ClnWffHW25PT4kII .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-ClnWffHW25PT4kII .icon-shape p,#mermaid-svg-ClnWffHW25PT4kII .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-ClnWffHW25PT4kII .icon-shape .label rect,#mermaid-svg-ClnWffHW25PT4kII .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-ClnWffHW25PT4kII .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-ClnWffHW25PT4kII .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-ClnWffHW25PT4kII :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
手工编码知识
从数据中学习
符号主义显式规则 + 知识库
智能
联结主义连接权重 + 分布式表示
联结主义为智能体赋予了强大的感知和模式识别能力，让机器可以直接从图像、声音和文本等原始数据中理解世界。但“看到世界”之后，智能体还需要学会“在动态环境中做序贯决策”。
5.2 强化学习：通过交互和试错学习决策
强化学习解决的是决策问题。它不是从标注好的静态数据集中学习，而是让智能体与环境直接交互，通过奖励信号不断优化长期收益。
#mermaid-svg-DlGqhuRHD4llmvHp{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-DlGqhuRHD4llmvHp .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-DlGqhuRHD4llmvHp .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-DlGqhuRHD4llmvHp .error-icon{fill:#552222;}#mermaid-svg-DlGqhuRHD4llmvHp .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-DlGqhuRHD4llmvHp .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-DlGqhuRHD4llmvHp .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-DlGqhuRHD4llmvHp .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-DlGqhuRHD4llmvHp .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-DlGqhuRHD4llmvHp .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-DlGqhuRHD4llmvHp .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-DlGqhuRHD4llmvHp .marker{fill:#333333;stroke:#333333;}#mermaid-svg-DlGqhuRHD4llmvHp .marker.cross{stroke:#333333;}#mermaid-svg-DlGqhuRHD4llmvHp svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-DlGqhuRHD4llmvHp p{margin:0;}#mermaid-svg-DlGqhuRHD4llmvHp .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-DlGqhuRHD4llmvHp .cluster-label text{fill:#333;}#mermaid-svg-DlGqhuRHD4llmvHp .cluster-label span{color:#333;}#mermaid-svg-DlGqhuRHD4llmvHp .cluster-label span p{background-color:transparent;}#mermaid-svg-DlGqhuRHD4llmvHp .label text,#mermaid-svg-DlGqhuRHD4llmvHp span{fill:#333;color:#333;}#mermaid-svg-DlGqhuRHD4llmvHp .node rect,#mermaid-svg-DlGqhuRHD4llmvHp .node circle,#mermaid-svg-DlGqhuRHD4llmvHp .node ellipse,#mermaid-svg-DlGqhuRHD4llmvHp .node polygon,#mermaid-svg-DlGqhuRHD4llmvHp .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-DlGqhuRHD4llmvHp .rough-node .label text,#mermaid-svg-DlGqhuRHD4llmvHp .node .label text,#mermaid-svg-DlGqhuRHD4llmvHp .image-shape .label,#mermaid-svg-DlGqhuRHD4llmvHp .icon-shape .label{text-anchor:middle;}#mermaid-svg-DlGqhuRHD4llmvHp .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-DlGqhuRHD4llmvHp .rough-node .label,#mermaid-svg-DlGqhuRHD4llmvHp .node .label,#mermaid-svg-DlGqhuRHD4llmvHp .image-shape .label,#mermaid-svg-DlGqhuRHD4llmvHp .icon-shape .label{text-align:center;}#mermaid-svg-DlGqhuRHD4llmvHp .node.clickable{cursor:pointer;}#mermaid-svg-DlGqhuRHD4llmvHp .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-DlGqhuRHD4llmvHp .arrowheadPath{fill:#333333;}#mermaid-svg-DlGqhuRHD4llmvHp .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-DlGqhuRHD4llmvHp .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-DlGqhuRHD4llmvHp .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-DlGqhuRHD4llmvHp .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-DlGqhuRHD4llmvHp .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-DlGqhuRHD4llmvHp .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-DlGqhuRHD4llmvHp .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-DlGqhuRHD4llmvHp .cluster text{fill:#333;}#mermaid-svg-DlGqhuRHD4llmvHp .cluster span{color:#333;}#mermaid-svg-DlGqhuRHD4llmvHp div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-DlGqhuRHD4llmvHp .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-DlGqhuRHD4llmvHp rect.text{fill:none;stroke-width:0;}#mermaid-svg-DlGqhuRHD4llmvHp .icon-shape,#mermaid-svg-DlGqhuRHD4llmvHp .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-DlGqhuRHD4llmvHp .icon-shape p,#mermaid-svg-DlGqhuRHD4llmvHp .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-DlGqhuRHD4llmvHp .icon-shape .label rect,#mermaid-svg-DlGqhuRHD4llmvHp .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-DlGqhuRHD4llmvHp .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-DlGqhuRHD4llmvHp .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-DlGqhuRHD4llmvHp :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
行动 A_t
新状态 S_t+1
奖励 R_t+1
智能体策略 π
环境
强化学习的核心要素包括：
- 智能体：学习者和决策者。
- 环境：智能体交互的外部世界。
- 状态：环境在某一时刻的描述。
- 行动：智能体根据状态采取的操作。
- 奖励：环境对行动好坏给出的标量反馈。
AlphaGo 就是典型例子。它观察棋盘布局，选择落子位置，在一局结束后根据胜负获得奖励，再通过大量自我对弈优化策略。学习目标不是最大化单步奖励，而是最大化从当前时刻到未来的累积回报。
5.3 预训练：让模型先拥有世界知识
强化学习通常需要大量针对特定任务的交互数据，智能体在学习之初缺乏先验知识。预训练的思路是：先让模型在互联网级通用语料上学习语言规律、事实知识和上下文逻辑，再针对下游任务进行微调。
#mermaid-svg-qqU0M8g6u4mvN6K8{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-qqU0M8g6u4mvN6K8 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-qqU0M8g6u4mvN6K8 .error-icon{fill:#552222;}#mermaid-svg-qqU0M8g6u4mvN6K8 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-qqU0M8g6u4mvN6K8 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-qqU0M8g6u4mvN6K8 .marker.cross{stroke:#333333;}#mermaid-svg-qqU0M8g6u4mvN6K8 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-qqU0M8g6u4mvN6K8 p{margin:0;}#mermaid-svg-qqU0M8g6u4mvN6K8 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-qqU0M8g6u4mvN6K8 .cluster-label text{fill:#333;}#mermaid-svg-qqU0M8g6u4mvN6K8 .cluster-label span{color:#333;}#mermaid-svg-qqU0M8g6u4mvN6K8 .cluster-label span p{background-color:transparent;}#mermaid-svg-qqU0M8g6u4mvN6K8 .label text,#mermaid-svg-qqU0M8g6u4mvN6K8 span{fill:#333;color:#333;}#mermaid-svg-qqU0M8g6u4mvN6K8 .node rect,#mermaid-svg-qqU0M8g6u4mvN6K8 .node circle,#mermaid-svg-qqU0M8g6u4mvN6K8 .node ellipse,#mermaid-svg-qqU0M8g6u4mvN6K8 .node polygon,#mermaid-svg-qqU0M8g6u4mvN6K8 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-qqU0M8g6u4mvN6K8 .rough-node .label text,#mermaid-svg-qqU0M8g6u4mvN6K8 .node .label text,#mermaid-svg-qqU0M8g6u4mvN6K8 .image-shape .label,#mermaid-svg-qqU0M8g6u4mvN6K8 .icon-shape .label{text-anchor:middle;}#mermaid-svg-qqU0M8g6u4mvN6K8 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-qqU0M8g6u4mvN6K8 .rough-node .label,#mermaid-svg-qqU0M8g6u4mvN6K8 .node .label,#mermaid-svg-qqU0M8g6u4mvN6K8 .image-shape .label,#mermaid-svg-qqU0M8g6u4mvN6K8 .icon-shape .label{text-align:center;}#mermaid-svg-qqU0M8g6u4mvN6K8 .node.clickable{cursor:pointer;}#mermaid-svg-qqU0M8g6u4mvN6K8 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-qqU0M8g6u4mvN6K8 .arrowheadPath{fill:#333333;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-qqU0M8g6u4mvN6K8 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-qqU0M8g6u4mvN6K8 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-qqU0M8g6u4mvN6K8 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-qqU0M8g6u4mvN6K8 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-qqU0M8g6u4mvN6K8 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-qqU0M8g6u4mvN6K8 .cluster text{fill:#333;}#mermaid-svg-qqU0M8g6u4mvN6K8 .cluster span{color:#333;}#mermaid-svg-qqU0M8g6u4mvN6K8 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-qqU0M8g6u4mvN6K8 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-qqU0M8g6u4mvN6K8 rect.text{fill:none;stroke-width:0;}#mermaid-svg-qqU0M8g6u4mvN6K8 .icon-shape,#mermaid-svg-qqU0M8g6u4mvN6K8 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-qqU0M8g6u4mvN6K8 .icon-shape p,#mermaid-svg-qqU0M8g6u4mvN6K8 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-qqU0M8g6u4mvN6K8 .icon-shape .label rect,#mermaid-svg-qqU0M8g6u4mvN6K8 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-qqU0M8g6u4mvN6K8 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-qqU0M8g6u4mvN6K8 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-qqU0M8g6u4mvN6K8 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
自监督预训练
任务数据微调
任务数据微调
任务数据微调
互联网级通用语料
基础大模型
情感分析
机器翻译
代码生成
当模型规模跨越某个阈值后，还会出现两种重要能力：
- 上下文学习：无需调整权重，只在输入中提供少量示例甚至零示例，模型就能完成新任务。
- 思维链推理：引导模型在回答复杂问题前先输出逐步推理，可以显著提升逻辑、算术和常识推理能力。
这意味着 LLM 不再只是一个语言模型，而是一个同时具备海量知识库和通用推理引擎双重角色的组件。
5.4 LLM 智能体：把历史技术拼图组合起来
到这一步，几块关键拼图已经出现：符号主义提供逻辑推理框架，联结主义和强化学习提供学习与决策能力，大型语言模型提供通过预训练获得的世界知识和通用推理能力。
#mermaid-svg-4T7lsipcvNSzKQEd{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-4T7lsipcvNSzKQEd .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-4T7lsipcvNSzKQEd .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-4T7lsipcvNSzKQEd .error-icon{fill:#552222;}#mermaid-svg-4T7lsipcvNSzKQEd .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-4T7lsipcvNSzKQEd .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-4T7lsipcvNSzKQEd .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-4T7lsipcvNSzKQEd .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-4T7lsipcvNSzKQEd .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-4T7lsipcvNSzKQEd .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-4T7lsipcvNSzKQEd .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-4T7lsipcvNSzKQEd .marker{fill:#333333;stroke:#333333;}#mermaid-svg-4T7lsipcvNSzKQEd .marker.cross{stroke:#333333;}#mermaid-svg-4T7lsipcvNSzKQEd svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-4T7lsipcvNSzKQEd p{margin:0;}#mermaid-svg-4T7lsipcvNSzKQEd .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-4T7lsipcvNSzKQEd .cluster-label text{fill:#333;}#mermaid-svg-4T7lsipcvNSzKQEd .cluster-label span{color:#333;}#mermaid-svg-4T7lsipcvNSzKQEd .cluster-label span p{background-color:transparent;}#mermaid-svg-4T7lsipcvNSzKQEd .label text,#mermaid-svg-4T7lsipcvNSzKQEd span{fill:#333;color:#333;}#mermaid-svg-4T7lsipcvNSzKQEd .node rect,#mermaid-svg-4T7lsipcvNSzKQEd .node circle,#mermaid-svg-4T7lsipcvNSzKQEd .node ellipse,#mermaid-svg-4T7lsipcvNSzKQEd .node polygon,#mermaid-svg-4T7lsipcvNSzKQEd .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-4T7lsipcvNSzKQEd .rough-node .label text,#mermaid-svg-4T7lsipcvNSzKQEd .node .label text,#mermaid-svg-4T7lsipcvNSzKQEd .image-shape .label,#mermaid-svg-4T7lsipcvNSzKQEd .icon-shape .label{text-anchor:middle;}#mermaid-svg-4T7lsipcvNSzKQEd .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-4T7lsipcvNSzKQEd .rough-node .label,#mermaid-svg-4T7lsipcvNSzKQEd .node .label,#mermaid-svg-4T7lsipcvNSzKQEd .image-shape .label,#mermaid-svg-4T7lsipcvNSzKQEd .icon-shape .label{text-align:center;}#mermaid-svg-4T7lsipcvNSzKQEd .node.clickable{cursor:pointer;}#mermaid-svg-4T7lsipcvNSzKQEd .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-4T7lsipcvNSzKQEd .arrowheadPath{fill:#333333;}#mermaid-svg-4T7lsipcvNSzKQEd .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-4T7lsipcvNSzKQEd .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-4T7lsipcvNSzKQEd .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-4T7lsipcvNSzKQEd .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-4T7lsipcvNSzKQEd .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-4T7lsipcvNSzKQEd .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-4T7lsipcvNSzKQEd .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-4T7lsipcvNSzKQEd .cluster text{fill:#333;}#mermaid-svg-4T7lsipcvNSzKQEd .cluster span{color:#333;}#mermaid-svg-4T7lsipcvNSzKQEd div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-4T7lsipcvNSzKQEd .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-4T7lsipcvNSzKQEd rect.text{fill:none;stroke-width:0;}#mermaid-svg-4T7lsipcvNSzKQEd .icon-shape,#mermaid-svg-4T7lsipcvNSzKQEd .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-4T7lsipcvNSzKQEd .icon-shape p,#mermaid-svg-4T7lsipcvNSzKQEd .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-4T7lsipcvNSzKQEd .icon-shape .label rect,#mermaid-svg-4T7lsipcvNSzKQEd .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-4T7lsipcvNSzKQEd .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-4T7lsipcvNSzKQEd .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-4T7lsipcvNSzKQEd :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
原始输入
Observation
Tool Call
新状态与工具结果
外部环境
感知模块
规划模块
大语言模型
记忆模块
执行模块
工具箱
LLM 智能体的工作循环可以拆成四步：
这也就是第一章中 Thought-Action-Observation 范式在更完整架构中的体现。
六、把发展史压缩成一条时间线
符号主义、联结主义和行为主义三大思想流派在几十年间交织、竞争并最终融合。下面的时间线适合作为整章的速览。
#mermaid-svg-aQLBlC0YOK97ivHa{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-aQLBlC0YOK97ivHa .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-aQLBlC0YOK97ivHa .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-aQLBlC0YOK97ivHa .error-icon{fill:#552222;}#mermaid-svg-aQLBlC0YOK97ivHa .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-aQLBlC0YOK97ivHa .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-aQLBlC0YOK97ivHa .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-aQLBlC0YOK97ivHa .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-aQLBlC0YOK97ivHa .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-aQLBlC0YOK97ivHa .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-aQLBlC0YOK97ivHa .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-aQLBlC0YOK97ivHa .marker{fill:#333333;stroke:#333333;}#mermaid-svg-aQLBlC0YOK97ivHa .marker.cross{stroke:#333333;}#mermaid-svg-aQLBlC0YOK97ivHa svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-aQLBlC0YOK97ivHa p{margin:0;}#mermaid-svg-aQLBlC0YOK97ivHa .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-aQLBlC0YOK97ivHa .cluster-label text{fill:#333;}#mermaid-svg-aQLBlC0YOK97ivHa .cluster-label span{color:#333;}#mermaid-svg-aQLBlC0YOK97ivHa .cluster-label span p{background-color:transparent;}#mermaid-svg-aQLBlC0YOK97ivHa .label text,#mermaid-svg-aQLBlC0YOK97ivHa span{fill:#333;color:#333;}#mermaid-svg-aQLBlC0YOK97ivHa .node rect,#mermaid-svg-aQLBlC0YOK97ivHa .node circle,#mermaid-svg-aQLBlC0YOK97ivHa .node ellipse,#mermaid-svg-aQLBlC0YOK97ivHa .node polygon,#mermaid-svg-aQLBlC0YOK97ivHa .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-aQLBlC0YOK97ivHa .rough-node .label text,#mermaid-svg-aQLBlC0YOK97ivHa .node .label text,#mermaid-svg-aQLBlC0YOK97ivHa .image-shape .label,#mermaid-svg-aQLBlC0YOK97ivHa .icon-shape .label{text-anchor:middle;}#mermaid-svg-aQLBlC0YOK97ivHa .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-aQLBlC0YOK97ivHa .rough-node .label,#mermaid-svg-aQLBlC0YOK97ivHa .node .label,#mermaid-svg-aQLBlC0YOK97ivHa .image-shape .label,#mermaid-svg-aQLBlC0YOK97ivHa .icon-shape .label{text-align:center;}#mermaid-svg-aQLBlC0YOK97ivHa .node.clickable{cursor:pointer;}#mermaid-svg-aQLBlC0YOK97ivHa .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-aQLBlC0YOK97ivHa .arrowheadPath{fill:#333333;}#mermaid-svg-aQLBlC0YOK97ivHa .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-aQLBlC0YOK97ivHa .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-aQLBlC0YOK97ivHa .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-aQLBlC0YOK97ivHa .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-aQLBlC0YOK97ivHa .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-aQLBlC0YOK97ivHa .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-aQLBlC0YOK97ivHa .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-aQLBlC0YOK97ivHa .cluster text{fill:#333;}#mermaid-svg-aQLBlC0YOK97ivHa .cluster span{color:#333;}#mermaid-svg-aQLBlC0YOK97ivHa div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-aQLBlC0YOK97ivHa .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-aQLBlC0YOK97ivHa rect.text{fill:none;stroke-width:0;}#mermaid-svg-aQLBlC0YOK97ivHa .icon-shape,#mermaid-svg-aQLBlC0YOK97ivHa .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-aQLBlC0YOK97ivHa .icon-shape p,#mermaid-svg-aQLBlC0YOK97ivHa .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-aQLBlC0YOK97ivHa .icon-shape .label rect,#mermaid-svg-aQLBlC0YOK97ivHa .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-aQLBlC0YOK97ivHa .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-aQLBlC0YOK97ivHa .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-aQLBlC0YOK97ivHa :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}
1950-1970符号主义：PSSH、专家系统、SHRDLU
1966ELIZA
1986心智社会
1980s联结主义复兴
2016AlphaGo 与强化学习
2017Transformer
2020sLLM 智能体
进入 21 世纪 20 年代，这些流派以前所未有的方式融合。GPT 系列模型本身是联结主义的产物，却成为执行符号推理、工具调用和规划决策的核心大脑，形成神经符号结合的现代智能体架构。
七、小结
智能体发展史不是简单的技术迭代，而是一场关于如何定义智能、获取知识和做出决策的思想变革。
- 符号主义用“知识 + 推理”模拟智能，但受限于知识获取瓶颈、常识缺失和系统脆弱性。
- ELIZA 用很少的代码展示了规则系统的表面智能，也暴露了它缺乏理解、没有记忆和难以扩展的问题。
- 心智社会把智能解释为大量简单过程协作涌现的结果，为多智能体系统提供了概念基础。
- 联结主义把知识从手工规则变成分布式权重，强化学习让智能体从交互中学习决策。
- 预训练让模型获得海量世界知识，LLM 智能体则把逻辑推理、学习决策和工具执行组合成现代闭环。
理解这条演进路径后，再进入大语言模型原理和现代智能体框架，会更容易看清每一项技术究竟在解决哪一段历史遗留问题。
评论前必须登录！
注册