In RL, a policy is a function or strategy that maps states to actions. Formally, π(s) specifies the action the agent will take when in state s. In trading, a policy could be a rule like “if RSI < 30, then buy”. The goal of RL is to learn an optimal policy that maximizes expected reward (e.g. trading profit). In practice, policies can be stochastic (giving probabilities for actions) or deterministic. Training often involves adjusting the policy to improve performance on historic price data, akin to tuning parameters of a strategy.

Back to Robots Glossary Index