A new reinforcement learning algorithm predicts not just what an agent will choose, but how long it will take to decide.
Researchers have derived MI-SARSA, an on-policy temporal-difference algorithm that adds mutual-information regularization to standard reinforcement learning. The model learns a marginal action prior and penalizes an agent for using state-specific information that deviates from that prior, effectively pricing the cognitive cost of paying close attention to context. Because that same information cost shapes both behavior and computation, the model generates trial-level predictions for reaction time in addition to the usual predictions about choices and rewards. Most reinforcement learning models are silent on timing; MI-SARSA is not.
That matters because standard reinforcement learning assumes unlimited computation, which humans and animals plainly do not have. MI-SARSA offers a way to connect the reward-maximizing logic of RL to the perception, attention, and working-memory limits that actually govern biological decision-making. It also produces a concrete, testable link between policy complexity and reaction time, something psychologists have long observed but rarely modeled from first principles.
The catch, per the paper's own results, is that there's no free lunch: dialing up the information penalty makes agents faster and more robust to a changing environment, but it also caps how much reward they can ultimately earn. Bounded rationality, it turns out, is a tradeoff, not a hack.