Scaling Cooperative Intelligence via Inverse Planning and Probabilistic Programming

Abstract

How can we build cooperative machines that model and understand human minds — machines that assist us with our goals, coordinate on plans, infer the intentions behind our words, and even learn our norms and values? This thesis presents a scalable model-based approach to building such systems via inverse planning and probabilistic programming. First, we introduce a probabilistic programming architecture that implements a Bayesian theory of mind. This architecture, Sequential Inverse Plan Search (SIPS), performs online inference of human goals and plans by inverting a Bayesian model of incremental human planning. By combining high-performance symbolic planners with sequential Monte Carlo (SMC) inference, SIPS achieves faster-than-real-time speed, while scaling to hundreds of possible goals, and remaining robust to human mistakes due to boundedly-rational planning. Second, we present Cooperative Language-guided Inverse Plan Search (CLIPS), a system that integrates SIPS with large language models (LLMs) to model communicative cooperation. By using LLMs as likelihood functions within probabilistic programs, CLIPS can infer human goals from ambiguous instructions, then provide uncertainty-aware assistance with much higher levels of reliability than LLMs can on their own. In addition, CLIPS can be used to infer the shared intentions of communicating agents from their actions and words. Third, we show how inverse planning can model the acquisition of social normativity, formalizing norm-guided societal behavior as a norm-augmented stochastic game (NSG). In NSGs, agents assume that society follows a shared set of social norms, and infer these norms from the actions of other agents. By doing so, agents can rapidly learn cooperative social norms using orders of magnitude less data than model-free approaches. Finally, we present advances in probabilistic programming infrastructure that have enabled architectures such as SIPS and CLIPS. Through interfaces for programmable SMC and probabilistic programming with LLMs, developers can readily compose modeling and inference subroutines when designing probabilistically coherent intelligent systems. Together, these innovations demonstrate the feasibility and scalability of rational AI engineering for cooperatively intelligent machines, while illuminating the computational and algorithmic foundations of human cooperative intelligence.