Introduction
In the late 1970s, the Automated Mathematician (AM) program garnered significant attention by "discovering" fundamental mathematical concepts, such as prime numbers and Goldbach’s conjecture, starting from basic set theory. However, AM eventually plateaued, unable to evolve the very heuristics that guided its search. Its successor, Eurisko, attempted to overcome this by treating heuristics as discoverable concepts themselves. This paper provides a critical "post-mortem" on these systems, revealing that their initial success was less about "smart" heuristics and more about a fortuitous "impedance match" between the Lisp programming language and the structure of mathematics.
The Mechanics of Discovery: Representation and Mutation
AM and Eurisko operated through a combination of frame-based representation and syntactic mutation. Concepts were structured as "frames" with various "slots" containing metadata and executable code.
The systems performed discovery by making small, "blind" changes to the Lisp code within a concept's Algorithm slot. For instance, changing an AND to an OR or removing a recursive call could yield a new, potentially interesting concept. A heuristic agenda managed these tasks, ranking discoveries based on their perceived "interestingness."
The "Lisp-Math Match" and the Complexity Barrier
The most profound insight of the retrospective is that AM’s success was largely "fortuitous." In Lisp, the code for a mathematical predicate is often extremely concise (3-5 lines). Because of this brevity, a random syntactic mutation has a high probability—a high "hit rate"—of producing a meaningful new mathematical concept. For example, by simply dropping a recursive call from the LIST-EQUAL algorithm, AM naturally derived the concept of "Same-Length."
However, this "hit rate" plummeted when the system attempted to evolve heuristics. Unlike simple mathematical predicates, a heuristic represented as a two-page Lisp program is highly complex. A random "point mutation" in such a large block of code is almost certain to result in "garbage" rather than an improved heuristic. This highlights a fundamental tension between form (syntax) and content (meaning): if the mapping between the two is not "natural" or "dense," automated discovery becomes nearly impossible.
From AM to Eurisko: Achieving Granularity
To address the complexity barrier, Eurisko moved away from "opaque lumps" of Lisp code. Instead, it employed a more granular representation where a single heuristic was decomposed into dozens of small, specialized slots (e.g., InputTerminals, OutputTerminals). This increased granularity restored the high "hit rate" for mutations, allowing Eurisko to successfully evolve domain-specific rules in fields like VLSI design and naval wargaming.
Despite these successes, the authors acknowledge the role of the human observer. We often attribute meaning to a program's output, "filling in the blanks" and crediting the system with a discovery because it triggers our own understanding of a known concept.
Conclusion: The Importance of Representation
The ultimate bottleneck in AI discovery is not just the search algorithm, but the relationship between knowledge acquisition and representation. A discovery system is only as effective as the "closeness" between its internal language and the domain it explores. While Eurisko achieved success through manual tuning of domain-specific slots, the next frontier remains getting systems to automatically extend their own vocabulary of representations—a challenge that continues to define autonomous AI research today.
Summary updated