ERAHBO: A Bayesian Optimization Method for Risk-Aware AutoRL
Researchers propose ERAHBO (Efficient Heteroscedastic Bayesian Optimization for Risk-Averse AutoRL), a Bayesian optimization method that models both mean and variance of reinforcement learning outcomes as functions of hyperparameter configurations. ERAHBO aims to identify hyperparameters achieving high average return while reducing variability across training runs. It improves sample efficiency via adaptive re-sampling rather than a fixed budget per hyperparameter. Empirical evaluations across diverse RL algorithms and environments show ERAHBO generally outperforms both risk-neutral and risk-averse baselines in sample efficiency for risk-averse returns.
Key facts
- ERAHBO models both mean and variance of RL outcomes as functions of hyperparameter configurations.
- It uses adaptive re-sampling instead of a fixed budget per hyperparameter.
- ERAHBO outperforms risk-neutral and risk-averse baselines in sample efficiency.
- The method is evaluated across diverse RL algorithms and environments.
- It aims to achieve high average return while reducing variability across training runs.
Entities
—