Online Bayesian Goal Inference

Yuer Tang, Tao Gao

UCLA · Tao Gao Lab · 2023–2024
Live Demo Paper (coming soon)

Abstract

We model and infer agent goals within grid-based environments using Bayesian inference and reinforcement learning. This project explores how observers can predict goals from observed actions using structured probabilistic reasoning. We develop a real-time online visualization system that demonstrates goal inference as agents navigate toward unknown targets.

System overview figure — coming soon

Figure 1: Overview of the Online Bayesian Goal Inference system. An agent navigates a grid world while a Bayesian observer updates goal posteriors in real time.

Method

The system consists of three main components:

  • MDP Solver: Uses Value Iteration with Bellman updates to pre-compute optimal value functions for multiple hypothetical goals across the grid environment.
  • Policy Extractor: Implements a gamma-discounted softmax policy with a temperature parameter for controlled action stochasticity, allowing the model to account for sub-optimal agent behavior.
  • Bayesian Inference Engine: Performs real-time posterior updates over possible goals based on observed agent actions, using the extracted policies as likelihood functions.
Architecture diagram — coming soon

Figure 2: System architecture showing the MDP solver, policy extractor, and Bayesian inference pipeline.

Results

Key achievements of this project include:

  • Foundational MDP solver using Value Iteration and Bellman updates for pre-computing optimal value functions across multiple hypothetical goals
  • Gamma-discounted softmax policy extractor with temperature parameter for controlled action stochasticity
  • Robust Bayesian goal inference mechanism for real-time posterior updates based on observed actions
  • Full-stack, low-latency online visualization system using Python (RL backend) and Node.js/Socket.io (frontend) for real-time demonstration

Interactive Demo

The full-stack visualization system demonstrates real-time goal inference as agents navigate grid environments. The backend runs reinforcement learning computations in Python, while the frontend provides an interactive visualization via Node.js and Socket.io.

Launch Demo

Citation

@article{tang2024goalinference,
  title     = {Online Bayesian Goal Inference in Grid Environments},
  author    = {Tang, Yuer and Gao, Tao},
  year      = {2024},
  institution = {UCLA}
}