Unsupervised Learning of Efficient Exploration: Pre-training Adaptive Policies via Self-Imposed Goals

The article presents ULEE, an unsupervised meta-learning method that enhances reinforcement learning agents' exploration and adaptation capabilities by allowing them to set and pursue self-imposed goals. ULEE integrates an in-context learner with an adversarial goal-generation strategy, optimizing for efficient multi-episode exploration within a meta-learning framework. Benchmark results on XLand-MiniGrid demonstrate that ULEE improves zero-shot and few-shot performance compared to learning from scratch and other pre-training methods, making it a valuable approach for practitioners aiming to accelerate learning in diverse and unknown task environments.