Hala Alshammari

تخصص الذكاء الاصطناعي · دفعة 2026

Connecting Minimal Agent Models: From Finite State Controllers to Tiny Recurrent Neural Networks

There are two established ways to write down a small, readable model of how an animal or a person makes repeated decisions. A Finite State Controller (FSC) stores what matters as a handful of labelled memory states and a lookup table. A tiny recurrent neural network (RNN) stores the same thing as a short list of numbers updated by trained weights. The two look unrelated, and this project asks how they connect. We fit both to the same synthetic win-stay/lose-shift data, where the correct rule is known in advance, and compare what each one holds inside. The central result is that matching behavior is not the same as learning the rule: on every trial the data contains, the two models are interchangeable, yet a probe that steps just outside the data shows the trained network never learned win-stay at all. It learned a close approximation that slowly drifts to the wrong answer when pushed. We then show, by solving for the network’s weights rather than training them, that a network can hold the rule exactly, down to a single hidden unit, and we identify why training misses that solution. The work turns two loose intuitions into precise statements: exactly how much memory a network needs to mirror an 𝑀-state controller, and where a softmax readout cannot reproduce a controller’s policy.

اقرأ التقرير النهائي (PDF)