DENDRAL 开启专家系统路线

未知有机物进入质谱仪后，并不会把自己的分子结构直接打印出来。电子轰击使分子裂解，仪器记录碎片离子的图样；研究者再从这些碎片反推原子如何连接。候选结构可能有成百上千种。DENDRAL 接手的是这段密集、重复、极依赖专业判断的排除工作。

项目的一条线可以追溯到火星。约书亚·莱德伯格从 1961 年起参与 NASA 的火星任务讨论，设想由计算机控制的质谱设备分析表面样本、寻找生命迹象。远程仪器一旦测到光谱，就需要自动或半自动地解释数据。到 1965 年，这套关于自动解释光谱的思考回到斯坦福的化学实验室，成为 DENDRAL 的基础。莱德伯格、爱德华·费根鲍姆、卡尔·杰拉西以及布鲁斯·布坎南等人，把问题限定在已知类别的有机化合物中——例如特定族的分子——而不是宣称覆盖全部化学空间。

程序先生成可能的分子结构（“生成”），再利用化学约束和仪器数据逐个排除（“测试”）。某些原子组合在化学上不成立，某些候选无法产生观察到的碎片，于是搜索树的一批枝条可以剪掉。DENDRAL 这个名字取自表示“树”的希腊词根 dendron。领域知识在树上发挥作用，不断缩小程序需要检查的范围。通用问题求解器在缺少领域知识时容易遭遇组合爆炸；DENDRAL 的回答是：把专家知道的约束写进搜索。质谱数据提供证据，化学稳定性与裂解规律提供剪刀；二者缺一，树就剪不下去。

困难集中在 if 与 then 之间：化学家平日未必会明说的判断。知识工程师通过长期访谈抽取启发式规则，再把经验判断转换成符号形式。每条有用的规则同时也是一次解释劳动的记录：专家停下来，把默认假设、例外和条件说清楚。代码能走多远，取决于专家愿意花多少时间修订条件、暴露习惯。项目史记载，全面运行后，DENDRAL 在特定任务上能比专家质谱分析师更快，并达到相近的准确度；它适合作为助手，不能由此变成化学家。知识不是从数据中“学”出来的，而是由专家与工程师共同整理进知识库。

树外的成本

DENDRAL 把人工智能从棋类等封闭环境带进真实的生物医学研究。它启发了 MYCIN（感染性疾病抗生素建议）、XCON（DEC 的计算机配置）等系统，也使“知识库”“推理规则”“知识获取”成为 1970–1980 年代一条产业路线。然而，美国国家医学图书馆保存的项目史同样明确：这类系统是定制品，需要多年高强度劳动和持续资助；十多年后，DENDRAL 仍只能处理若干特定化合物类别，不能覆盖全部化学空间。规则维护昂贵、例外难穷尽、环境一变知识就过时——这些问题在商业热潮中会被放大。

它所依赖的计算环境也在扩大。1973 年启用的 SUMEX-AIM 把斯坦福的医学 AI 计算资源通过 ARPANET 提供给外地研究者；到 1980 年，平台托管了 19 个项目，其中包括 DENDRAL 的后继系统、MYCIN、MOLGEN 和匹兹堡大学的 INTERNIST。专家系统由此不再只是某个实验室的一段程序，也成了一套共享机器、网络连接、领域专家和长期资助共同维持的研究设施。

费根鲍姆后来常把这类工作概括为“知识就是力量”：在组合爆炸的搜索中，领域约束比更聪明的通用搜索更能缩小空间。DENDRAL 的实践给这句话加了脚注——知识确实有力量，但前提是有人持续支付抽取、形式化与修订的成本。MYCIN 把类似方法用于感染性疾病的抗生素选择，并引入确定性因子等不确定推理表示；XCON 则把规则推到计算机配置的生产线旁。成功案例加速了 1980 年代的商业热潮，也把维护负债一起放大。

写进规则的知识扩大了专业判断的作用范围；没有写出来、无法稳定表述、会随环境改变的部分，则留在系统边界之外。DENDRAL 的树剪去了大量候选，也逐枝显示出知识工程能够抵达哪里——以及抵达之后，还要付多少维护的账。质谱碎片仍由仪器产生；解释它们的规则，却必须由人反复重写。

A mass spectrometer does not print the structure of an unknown organic molecule. It records a pattern of fragment ions produced when the molecule breaks apart under electron bombardment. From those pieces, a chemist must infer how the original atoms were connected. Hundreds or thousands of candidate structures may fit the gross formula. DENDRAL was built for the long work of generating, testing, and eliminating them.

One strand of the project began with Mars. From 1961, Joshua Lederberg participated in NASA discussions of instruments that might analyze surface samples for signs of life. The prospect of a computer-controlled mass spectrometer raised a second problem: an instrument in a remote environment would need help interpreting what it measured. By 1965, this interest in automated spectrum analysis had returned to the terrestrial chemistry laboratory at Stanford. Lederberg, Edward Feigenbaum, Carl Djerassi, Bruce Buchanan, and their collaborators deliberately bounded the task to known classes of organic compounds—selected families of molecules—rather than claiming to cover all of chemistry.

DENDRAL generated possible molecular structures, then pruned them with chemical constraints and evidence from the spectrum. Some atomic arrangements were chemically impossible. Some candidates could not have produced the observed fragments. Entire branches of the search could be discarded at once. The name came from a Greek root for tree (dendron), an unusually literal description of where expertise entered the computation. General problem solvers without domain knowledge met combinatorial explosion; DENDRAL’s reply was to write the chemist’s constraints into the search. Mass spectra supplied evidence; chemical stability and fragmentation regularities supplied the shears. Without both, the tree could not be cut.

The hard part lay between if and then. A working chemist’s judgment is full of unstated conditions, defaults, and exceptions. Knowledge engineers conducted extended interviews to extract these heuristics and convert them into symbolic rules. Each useful rule was therefore also a record of explanation: an expert stopping to articulate what routine practice had allowed them to leave unsaid. The system’s reach depended on continued revision of that record. Within selected tasks, project histories report that DENDRAL could work faster than expert mass-spectrometry analysts while reaching comparable accuracy. That made it a serious scientific assistant, not an electronic chemist. Its knowledge was engineered with experts rather than learned from data.

The approach later reappeared in MYCIN (antibiotic advice for infectious disease), XCON (computer configuration for Digital Equipment Corporation), and a commercial expert-systems industry built around knowledge bases, inference rules, and knowledge acquisition. The archived project history keeps the cost visible. DENDRAL was custom-built, required years of intensive labor and sustained funding, and after more than a decade still handled only certain compound classes rather than all of chemistry. Explicit knowledge travels well through software; tacit, shifting, or poorly expressible knowledge does not become explicit merely because the product is called an expert system. Rule maintenance, conflicting exceptions, and environments that refuse to stay fixed would later become commercial liabilities.

The computing environment around the work expanded as well. SUMEX-AIM, established in 1973, made Stanford’s medical-AI resources available to remote researchers over the ARPANET. By 1980 it hosted nineteen projects, among them DENDRAL and its descendants, MYCIN, MOLGEN, and the University of Pittsburgh’s INTERNIST. An expert system was no longer only a program in one laboratory. It was also an institution made of shared machines, network access, domain specialists, and continuing public support.

Feigenbaum and others later summarized the line of work with the slogan that “knowledge is power”: in a combinatorial search, domain constraints shrink the space more effectively than a cleverer general search alone. DENDRAL’s practice adds a footnote—knowledge has power only if someone keeps paying for extraction, formalization, and revision. MYCIN applied similar methods to antibiotic selection in infectious disease and introduced certainty factors as one representation of uncertain inference. XCON carried rules onto the factory floor of computer configuration. Those successes accelerated the commercial boom of the 1980s and scaled the maintenance liability with them.

DENDRAL’s tree lost vast numbers of candidates as chemistry cut away impossible branches. It also displayed, branch by branch, where the encoded knowledge ended—and what maintenance debt began once those branches were written down. The mass spectrometer still produced the fragments. The rules that explained them had to be rewritten by people, again and again.