Policy Compilation for Stochastic Constraint Programs

Abstract

Real-world risk-bounded planning and decision-making problems are fluid, uncertain, and highly dynamic, demanding an architecture which can encode and solve a rich set of problems involving decision-making under uncertainty. While many solution architectures exist for solving deterministic CSPs, very few are able to generate decisions that are robust to uncontrolled, stochastic events, and even fewer are able to construct conditional policies that are able to adapt online to these uncertain outcomes. In this thesis, I present a variant of the Optimal Satisfiability Problem Solver (OpSat) that solves dynamic, chance-constrained satisfiability problems. The proposed variant solves these real-world problems efficiently and encodes policies compactly through a hybrid architecture that (a) encodes probabilistic information explicitly as logical constraints, (b) performs temporal reasoning to extract logical temporal conflicts, and (c) compiles out the constraints of a Weighted, Conditional, Stochastic CSP into a compact policy representation which may be efficiently queried. Such an architecture facilitates the design of robust, risk-aware systems by providing a user with the ability to solve a rich set of problems involving mixed logical and temporal constraints.¹

¹This research was generously supported by Airbus SE.