Domain-Specific Probabilistic Programming with Multiverse Explorer

Abstract

We present Multiverse Explorer, a domain-specific probabilistic programming language presented as a visual language integrated with a domain world model. The interactive visualisation presents a Monte Carlo simulation over a causal graph, allowing the user to gain an overview and query alternative outcomes in a counterfactual manner. Separate graphs express the policies attributed to multiple heterogeneous agents. The outcomes of actions are visualised in an interactive 3D animation of the environment; in this work, we apply the Multiverse Explorer to multi-agent driving scenarios by extending the CARLA simulator. The Multiverse Explorer has been evaluated with a sample of technical non-specialists, demonstrating the potential of this approach to be used in design, audit, policy, litigation, and other contexts where the outcome of multi-agent decision scenarios must be investigated by professionals beyond a specialist AI audience.