ATLAS: A New Continual Learning Algorithm for Reinforcement Learning
A novel algorithm named Adaptive Topological Learning with Abstract Successors (ATLAS) has been developed to tackle issues related to sample efficiency and robustness in reinforcement learning. The research, published on arXiv (2608.04334), introduces a technique that merges a Grow When Required network with Successor Features, enhancing sample efficiency and reducing catastrophic forgetting. ATLAS has been tested in spatial navigation scenarios, compared with standard on-policy and off-policy algorithms. Findings indicate that by separating transition dynamics from the reward signal, ATLAS can quickly adapt to new objectives and demonstrates beneficial backward transfer, significantly surpassing traditional methods in dynamic environments.
Key facts
- ATLAS stands for Adaptive Topological Learning with Abstract Successors.
- The paper is available on arXiv with identifier 2608.04334.
- ATLAS uses a Grow When Required network with Successor Features.
- It aims to achieve high sample efficiency and robustness to environmental changes.
- Evaluated in spatial navigation tasks.
- Benchmarked against on-policy and off-policy algorithms.
- ATLAS decouples transition dynamics from the reward signal.
- It exhibits positive backward transfer and near-instantaneous adaptation to new goals.
Entities
Institutions
- arXiv