Meta-reinforcement Learning Task Planning Based on Improved Curriculum Learning Sampling

Model-Agnostic Meta-Learning (MAML), have shown a good generalization performance in solving task planning problems. Most of the existing meta-reinforcement learning(meta-rl) algorithms learn meta-policies by randomly sampling tasks. However, this sampling strategy can only adapt to simple tasks, and have insufficient ability to generalize to difficult and new tasks. To address these problems, the Improved curriculum learning-MAML (ICL-MAML) framework is designed, which the task sampling strategy based on improved curriculum learning is applied for sampling the tasks. This method improves the convergence ability of meta strategies for difficult tasks and their generalization ability for new tasks. At the same time, it alleviates the problem of catastrophic forgetting. Experiments in the meta-world environment show that ICL-MAML has a significant improvement in training and testing success rates, and uses less training time compared to random sampling task strategy.