Goal Recognition Design for General Behavioral Agents using Machine Learning

This repository contains the code and data to reproduce results for the paper "Goal Recognition Design for General Behavioral Agents using Machine Learning". We provide code for modifying environment designs to mininimize the Worst-Case Distinctiveness (WCD) using our Lagrange-based Gradient Descent approach. The provided code covers the different behavioral settings in our paper: optimal agents, suboptimal agents with hyperbolic discounting, and human behavior data.

Overview

The repository provides:

• Simulated data generation for environments with corresponding WCD values
• WCD prediction model training scripts
• Optimization code using trained WCD predictive models to modify environments
• Baseline implementations including Greedy, Exhaustive Search, and Pruned-Reduce methods
• Multiple behavioral settings: optimal agents, suboptimal agents, and human behavior data
• Overcooked-AI domain extension with specific implementations

Prerequisites

• Python: Python 3.10.6 or higher
• CUDA: GPU support recommended for faster training and optimization
• Dependencies: Install packages from environment.yml using conda

Repository Structure

├── optimal/                    # Optimal agent experiments
│   ├── README.md              # Detailed guide for optimal setting
│   ├── train_wcd_predictor_optimal.py # WCD prediction model training
│   ├── run_optimization_optimal.py # Main optimization experiments
│   ├── baselines/             # Baseline experiment implementations
│   └── models/                # Trained models for grid sizes 6 and 13
│
├── suboptimal/                # Suboptimal agent experiments
│   ├── README.md              # Detailed guide for suboptimal setting
│   ├── train_wcd_predictor_suboptimal.py # WCD prediction model training
│   ├── run_optimization_suboptimal.py # Main optimization experiments
│   ├── simulate_training_data_suboptimal.py # Training data generation
│   └── models/                # Trained models for grid size 6
│
├── human-exp-data-driven/     # Human behavior experiments
│   ├── README.md              # Detailed guide for human behavior setting
│   ├── run_optimization_human_bhvr.py # Main optimization experiments
│   └── baselines/             # Baseline experiment implementations
│
├── overcooked-ai/             # Overcooked-AI domain extension
│   ├── README.md              # Detailed guide for Overcooked-AI
│   └── src/                   # Implementation code
│
└── human-behavior-data/       # MTurk human subject experiment data

Experimental Settings

Optimal Agents: For environments with optimal agent behavior

• Use case: Standard gridworld optimization with optimal pathfinding
• Key features: Multiple model architectures (CNN, Transformer, GNN, etc.)
• See: optimal/README.md for detailed instructions

Suboptimal Agents: For environments with hyperbolic discounting agents

• Use case: Realistic agent behavior with time discounting
• Key features: K parameter controls suboptimality level
• See: suboptimal/README.md for detailed instructions

Human Behavior: For data-driven human behavior experiments

• Use case: Real human behavior from MTurk experiments
• Key features: Uses actual human trajectory data
• See: human-exp-data-driven/README.md for detailed instructions

Overcooked-AI: For the Overcooked-AI domain

• Use case: Multi-agent coordination scenarios
• Key features: Domain-specific implementations
• See: overcooked-ai/README.md for detailed instructions

Install dependencies

Install the required packages using the repository’s environment.yml:

conda env create -f environment.yml
conda activate data-driven-grd

Regenerate Paper Figures

Use the unified figure driver from the repository root:

python manuscript_figures/generate_all.py --list

Generate specific figures:

python manuscript_figures/generate_all.py 3a 3b 4a 4b 5a 5b 6a 6b 8a 8b 8c 8d 9a 9b

Generate all available figure drivers:

python manuscript_figures/generate_all.py

Generated outputs are saved in manuscript_figures/generated_figures/.

Notes:

• Use an environment where python has the project dependencies (especially PyTorch) if you want full recomputation from raw data.

Citation

If you use this framework in your research, please cite the following.

@article{kasumba2025data,
  title={Data-driven goal recognition design for general behavioral agents},
  author={Kasumba, Robert and Yu, Guangyu and Ho, Chien-Ju and Keren, Sarah and Yeoh, William},
  journal={Transactions in Machine Learning Research},
  year={2025}
}