Beyond Single Policies: TeLAPA Preserves Plasticity via Latent-Aligned Archives in Continual RL
TeLAPA organizes diverse policy archives in a shared latent space to retain plasticity and accelerate relearning in continual RL, outperforming single-model preservation on MiniGrid tasks.
A new framework called TeLAPA maintains archives of behaviorally diverse policies in a shared latent space to counter loss of plasticity after task interference in continual reinforcement learning (Lillo, arXiv:2604.15414, 2026).
TeLAPA draws on quality-diversity optimization to build per-task policy neighborhoods rather than committing to one evolving model, demonstrating higher task success rates, faster recovery on revisited tasks, and stronger retention in MiniGrid continual RL benchmarks (Lillo, arXiv:2604.15414, 2026). Analyses within the work establish that source-optimal policies frequently differ from transfer-optimal ones even inside competent local neighborhoods, showing that collapsing archives to a single representative discards useful plasticity (Lillo, arXiv:2604.15414, 2026; Mouret & Clune, arXiv:1504.04909, 2015).
Standard single-model techniques such as elastic weight consolidation address catastrophic forgetting yet overlook behavioral diversity needed for rapid re-adaptation, a gap visible when comparing against MAP-Elites style illumination of search spaces and recent continual RL plasticity studies (Kirkpatrick et al., PNAS 2017; Dohare et al., arXiv:2305.16211, 2024). The original TeLAPA abstract understates how preserving skill-aligned neighborhoods reframes continual RL away from isolated solution retention toward reusable policy collections under distribution drift.
Preservation of multiple nearby alternatives in latent space directly mitigates the plasticity barrier repeatedly observed across deep network continual learning trajectories, supplying a concrete mechanism that connects MiniGrid results to open-ended lifelong agent architectures required for AGI-scale adaptation (Lillo, arXiv:2604.15414, 2026; Kirkpatrick et al., PNAS 2017).
AXIOM: Single-policy optimization in continual RL collapses neighborhoods that preserve plasticity; maintaining latent-aligned archives of diverse competent policies enables faster recovery and may remove a core obstacle to scalable lifelong learning agents.
Sources (3)
- [1]Beyond Single-Model Optimization: Preserving Plasticity in Continual Reinforcement Learning(https://arxiv.org/abs/2604.15414)
- [2]Overcoming Catastrophic Forgetting in Neural Networks(https://www.pnas.org/doi/10.1073/pnas.1611835114)
- [3]MAP-Elites: Illuminating Search Spaces(https://arxiv.org/abs/1504.04909)