Efficient Hyperparameter Optimization for Physics-based Character Animation

Physics-based character animation has seen significant advances in recent years with the adoption of Deep Reinforcement Learning (DRL). However, DRL-based learning methods are usually computationally expensive and their performance crucially depends on choice hyperparameters. Tuning hyperparameters for these often requires repetitive training control policies, which is even more prohibitive. In this work, we propose a novel Curriculum-based Multi-Fidelity Bayesian Optimization framework (CMFBO) efficient hyperparameter optimization systems. Using curriculum-based task difficulty as fidelity criterion, our method improves searching efficiency by gradually pruning search space through evaluation easier motor skill tasks. We evaluate two physics-based tasks: morphology tuning DeepMimic. Our algorithm significantly outperforms state-of-the-art applicable animation. particular, show that optimized result at least 5x gain comparing to author-released settings