Balancing Plasticity and Stability with Fast and Slow Successor Features

In three linesStudy on RL agent adaptation in gradually non-stationary environments. Authors modify 3D Miniworld and MuJoCo environments to introduce continuous drift, showing that synaptic consolidation applied to multi-timescale Successor Features outperforms Q-value-based approaches. Stability outweighs plasticity in continual learning with gradual changes.

Summary generated by Claude — human-verified