Hyperparameter Auto-Tuning in Self-Supervised Robotic Learning

Policy optimization in reinforcement learning requires the selection of numerous hyperparameters across different environments. Fixing them incorrectly may negatively impact performance leading notably to insufficient or redundant learning. Insufficient (due convergence local optima) results under-performing policies whilst wastes time and resources. The effects are further exacerbated when using single solve multi-task problems. Observing that Evidence Lower Bound (ELBO) used Variational Auto-Encoders correlates with diversity image samples, we propose an auto-tuning technique based on ELBO for self-supervised Our approach can auto-tune three hyperparameters: replay buffer size, number policy gradient updates during each epoch, exploration steps epoch. We use a state-of-the-art robot framework (Reinforcement Learning Imagined Goals (RIG) Soft Actor-Critic) as baseline experimental verification. Experiments show our method online yields best at fraction computational Code, video, appendix simulated real-robot experiments be found project page www.JuanRojas.net/autotune.