Habits Have Geometry: Action Representation as Morphological Computation

Bounded organisms face evolutionary pressure to act cheaply: every decision carries an informational cost, and one under- explored way to reduce that cost is to reshape the represen- tation of action itself. We study action twists: given a world with state transitions induced by actions, we introduce a per- state relabelling of actions. In the “twisted” worlds the same state transitions are available as in the original, but the ac- tion labels corresponding to these may change from state to state. A genetic algorithm now searches for twists in grid- world environments that reduce decision information across many goals simultaneously. The discovered relabellings align action labels with environmental structure, exploiting walls and doorways as a form of morphological computation, and the benefit persists across a wide range of bounded-rationality regimes. Even on a featureless torus where no environmen- tal structure exists to exploit, the GA constructs habit cycles: short, label-specific orbits that the agent enters by repeating a single action symbol. Each habit cycle gives the agent a coherent cheap path across the environment, supporting our claim that action representation is itself a form of morpholog- ical computation, with its own geometry that can be evolved.