Dynamic Successor Features for transfer learning and guided exploration

摘要

The Successor Feature framework for Reinforcement Learning algorithms improves task transfer by decomposing the learned state–action value function. The decomposition involves two components, one that captures future-expected state features and the other that models the task-related reward structure. However, successful transfer between tasks depends heavily on how the reward function changes, possibly leading to failure of the original Successor Feature formulation. This paper proposes the Dynamic Successor Feature framework, DynSF, by extending the mathematical formulation of the original Successor Feature framework to center around a learned state-transition model. Under this formulation, the state-transition model dynamically induces the acting policy. The flexibility of DynSF also extends to the architecture, requiring only a state-transition model and a small vector of parameters. This architecture provides immense flexibility in the choice of the model used to learn the state-transition model. The DynSF framework is evaluated and compared to other baseline algorithms through several experiments in a continuous grid world environment, a robotic Reacher, and pixels in the Doom environment.