Meta-Active Learning in Probabilistically Safe Optimization

When a robotic system is faced with uncertainty, the must take calculated risks to gain information as efficiently possible while ensuring safety. The need safely and in face of uncertainty spans domains from healthcare search rescue. To learn when data scarce or difficult label, active learning acquisition functions intelligently select point that, if label were known, would most improve estimate unknown model. Unfortunately, prior work suffers an inability accurately quantify information-gain, generalize new domains, ensure safe operation. overcome these limitations, we develop Safe MetAL, probabilistically-safe, algorithm which meta-learns function for selecting sample efficient points safety critical domains. key our approach novel integration meta-active chance-constrained optimization. We (1) meta-learn based on history, (2) encode this optimization framework, (3) solve information-rich set enforcing probabilistic guarantees. present state-of-the-art results model damaged UAV optimal parameters deep brain stimulation. Our achieves 41% improvement 20% speedup computation time compared meta-learning approaches system.