ATLAS Boosts Continual Reinforcement Learning with Adaptive Topology

Traditional model-free reinforcement learning (RL) algorithms often suffer from low sample efficiency and a lack of robustness when environments change, leading to "catastrophic forgetting." Model-based RL improves sample efficiency but still struggles with environmental shifts. This paper introduces ATLAS, an Adaptive Topological Learning with Abstract Successors framework, to overcome these limitations. ATLAS employs a Grow When Required network combined with Successor Features. This architecture allows the algorithm to structurally decouple transition dynamics from the reward signal, a key innovation for robust continual learning. By doing so, ATLAS can adapt to new goals almost instantly and exhibits positive backward transfer, meaning it can improve performance on previously learned tasks even after learning new ones. Evaluated in spatial navigation tasks, ATLAS significantly outperforms common on-policy and off-policy baseline algorithms in non-stationary environments. Its ability to maintain high sample efficiency while effectively combating catastrophic forgetting represents a significant step forward for RL systems operating in dynamic and evolving real-world scenarios.