Behavior-Function-Guided Exploration in Deep Q-Learning
A deep reinforcement learning exploration method that augments a Deep Q-Network with a separately learned behavior function estimating how often the data-collection policy has taken each action in each state. This function is combined with the Q-function to generate a diverse population of policies that trade off state-action coverage, correction of value overestimation, and exploitation, with a lightweight adaptive meta-controller choosing one policy per episode. It improves exploration generality with minimal computational overhead and no task-specific inductive bias.
𝛽-DQN: Improving Deep Q-Learning By Evolving the Behavior Hongming Zhang
Beta-DQN augments standard Deep Q-Networks with a learned behavior function beta that estimates, for each state, the probability the data-collection policy took each action. Combining beta with the Q…