[2604.00686] Full-Gradient Successor Feature Representations

Computer Science > Machine Learning arXiv:2604.00686 (cs) [Submitted on 1 Apr 2026] Title:Full-Gradient Successor Feature Representations Authors:Ritish Shrirao, Aditya Priyadarshi, Raghuram Bharadwaj Diddigi View a PDF of the paper titled Full-Gradient Successor Feature Representations, by Ritish Shrirao and 2 other authors View PDF HTML (experimental) Abstract:Successor Features (SF) combined with Generalized Policy Improvement (GPI) provide a robust framework for transfer learning in Reinforcement Learning (RL) by decoupling environment dynamics from reward functions. However, standard SF learning methods typically rely on semi-gradient Temporal Difference (TD) updates. When combined with non-linear function approximation, semi-gradient methods lack robust convergence guarantees and can lead to instability, particularly in the multi-task setting where accurate feature estimation is critical for effective GPI. Inspired by Full Gradient DQN, we propose Full-Gradient Successor Feature Representations Q-Learning (FG-SFRQL), an algorithm that optimizes the successor features by minimizing the full Mean Squared Bellman Error. Unlike standard approaches, our method computes gradients with respect to parameters in both the online and target networks. We provide a theoretical proof of almost-sure convergence for FG-SFRQL and demonstrate empirically that minimizing the full residual leads to superior sample efficiency and transfer performance compared to semi-gradient baselines i...