Stein Variational Goal Generation for adaptive Exploration in Multi-Goal Reinforcement Learning
This addresses the challenge of training agents in multi-goal RL with sparse rewards, offering an incremental improvement over existing methods for better generalization.
The paper tackled the problem of sparse rewards and discontinuities in multi-goal reinforcement learning by proposing Stein Variational Goal Generation (SVGG) to sample goals of intermediate difficulty, resulting in improved success coverage in hard exploration problems and demonstrating recovery in changing environments.
In multi-goal Reinforcement Learning, an agent can share experience between related training tasks, resulting in better generalization for new tasks at test time. However, when the goal space has discontinuities and the reward is sparse, a majority of goals are difficult to reach. In this context, a curriculum over goals helps agents learn by adapting training tasks to their current capabilities. In this work we propose Stein Variational Goal Generation (SVGG), which samples goals of intermediate difficulty for the agent, by leveraging a learned predictive model of its goal reaching capabilities. The distribution of goals is modeled with particles that are attracted in areas of appropriate difficulty using Stein Variational Gradient Descent. We show that SVGG outperforms state-of-the-art multi-goal Reinforcement Learning methods in terms of success coverage in hard exploration problems, and demonstrate that it is endowed with a useful recovery property when the environment changes.