强化学习大本营新作：如何破解「学新忘旧」困局

TL;DR - University of Alberta researchers propose FAME, a dual-system framework that formulates continual reinforcement learning as minimizing catastrophic forgetting while preserving fast adaptation. It matters for agents and robots that must learn new tasks after deployment without losing previously acquired skills.

• FAME defines task distance through differences in optimal value functions or policies and measures forgetting using prior-task state visitation distributions.
• A Fast Learner uses hypothesis testing to choose among meta-learned initialization, fine-tuning the previous policy, or training from scratch for each new task.
• A Meta Learner consolidates new knowledge by minimizing a formal forgetting objective, expressed through incremental maximum likelihood, KL divergence, or Wasserstein distance depending on the setting.
• Across MinAtar, Atari, and Meta-World experiments with DQN, PPO, and SAC, FAME reportedly outperformed several continual-learning baselines on average performance, forward transfer, and forgetting.