Reducing the Time Complexity of Goal-Independent Reinforcement Learning

Abstract

Concurrent Q-Learning (CQL) is a goal independent reinforcement learning technique that learns the action values to all states simultaneously. These action values may then be used in a similar way to eligibility traces to allow many action values to be updated at each time step. CQL learns faster than conventional Q-learning techniques with the added benefit of being able to apply all experiences gained performing one task to any new task within the problem domain. Unfortunately the update time complexity of CQL is O(|S|2x|A|). This paper presents a technique for reducing the update complexity of CQL to O(|A|) with little impact on performance.