Dynamic Choice Models of Human Decision-Making
Predicting and understanding human behavior is key for the safety and reliability of autonomous agents. Human decisions are driven by preferences and utility functions. These are typically unknown and need be inferred from observed behavior. Existing approaches have a number of limitations when applied to real world settings, which are characterized by large dimensionality, complex dynamics, and noisy observational data. The goal of this proposal is to develop a new family of probabilistic models of human decision-making to overcome these challenges. The models will be useful both for activity forecasting (e.g., predict if a pedestrian will cross) and for imitating demonstrated behavior (e.g., defensive driving).
The main research goals are:
- Increase the expressive power of existing models, e.g., incorporating non-linear utility functions
- Scale to very large and complex state spaces, e.g., with new approximate inference and planning schemes
- Develop scalable learning techniques for large, noisy datasets
- Provide new models and algorithms to learn from suboptimal decisions and multi-agent scenarios