通用问题求解器 GPS

1957 年，纽厄尔与西蒙已经用逻辑理论家证明机器能证明数学定理。但那在他们看来只是开始——他们真正想要的，是一个不针对任何具体领域的「通用问题求解器」。这个野心，直接体现在它的名字里：General Problem Solver，简称 GPS。

GPS 的工作方式后来被反复讲成教科书里的经典。它用「手段-目的分析」来解题：先看清楚自己现在在哪里、目标在哪里，找出两者之间的差距，然后选择能缩小差距的动作；如果一步到位做不到，就把目标拆成更小的子目标，一步步逼近。这种思路听上去很「人类」，因为它本来就是从人类的解题行为中提炼出来的。

这套方法在今天看来朴素，但它第一次把「智能」变成了一套可编程的通用流程——状态、目标、差距、算子。它不需要懂具体的领域，只靠通用的搜索策略。正是这个框架，奠定了后来整个符号主义 AI 的底层范式，也直接影响了专家系统和自动规划系统的设计。

但 GPS 也暴露了那条贯穿 AI 史的理想与现实的裂缝。它声称「通用」，实际解决问题的能力却相当有限：一旦问题稍稍复杂，搜索空间就爆炸，程序常常在原地打转。纽厄尔和西蒙对它的早期乐观，很快被现实修正，后来两人自己也承认，「通用」远比想象中困难。

回看 GPS，它的意义不在于当时解决过多少实际问题，而在于它提出的问题——机器能不能拥有不限于单一领域的智能。这个问题此后六十年一直被追问，从专家系统到今天的通用大模型，都是对 1957 年那个理想的不同回答。GPS 是这条长路上第一面迎风展开的旗帜，即使它没能走远，方向却被它指了出来。

By 1957 Newell and Simon had already used the Logic Theorist to prove that machines could prove theorems. But for them that was only the beginning—what they truly wanted was a problem solver that was not tied to any particular domain. That ambition was written into its name: the General Problem Solver, GPS.

GPS's method would later become a textbook classic. It solved problems through means–ends analysis: first see where you are and where you want to be, find the gap between them, then pick actions that shrink the gap; if one step is not enough, break the goal into smaller subgoals and approach it incrementally. The idea sounds strikingly "human," because it was distilled from observing how humans solve problems.

The approach seems simple today, but it was the first time "intelligence" had been turned into a programmable general procedure—state, goal, gap, operator. GPS needed no domain expertise, only generic search strategies. This framework became the underlying paradigm of all symbolic AI, and it directly shaped expert systems and automated planning.

But GPS also exposed a fault line that runs through all of AI history: the gap between ideal and reality. It claimed to be "general," yet its actual solving power was quite limited. As soon as a problem grew even slightly complex, the search space exploded and the program often spun in place. The early optimism of Newell and Simon was quickly revised by reality, and both later acknowledged that "generality" was far harder than it sounded.

Looking back, GPS's meaning lies less in the problems it actually solved and more in the question it raised—whether a machine could possess intelligence not confined to a single domain. That question has been asked ever since, and every answer, from expert systems to today's general-purpose models, is a different response to the 1957 ideal. GPS was the first banner raised on that long road; even though it did not travel far, it pointed the way.