On-line planning for constrained autonomous agents in an uncertain world

This project aims to develop admissible heuristics for constrained stochastic planning problems and integrate them into state-of-the-art on-line algorithms. AI systems, such as self-driving cars or energy management systems, make intelligent decisions to act near-optimally in uncertain environments, leading to savings, for example in energy use. But we also want assurances that AI systems will obey safety constraints. Solving constrained stochastic planning problems is key to building AI that is both robust and safe. Current heuristics for this problem are based on reductions to the deterministic single-objective case, making them weak. This project will explore ways to account for uncertainty and constraints in heuristic estimation.