# Achievement

## New model of reinforcement learning

### Project

IGERT: Interdisciplinary Training in Perceptual Science

### University

Rutgers University New Brunswick

(New Brunswick, NJ)

### PI

## Research Achievements

### New model of reinforcement learning

Reinforcement learning is the study of how human or virtual agents learn to select actions in order to maximize long-term performance. Most models are based on finite size (discrete) actions. IGERT trainees Ari Weinstein and Chris Mansley (advisor M. Littman) studied more realistic actions that fill some continuous space. The developed a new learning algorithm to create an action-planner that could handle both large state spaces and continuous action spaces natively. The algorithm generally outperforms discrete-action models even when the set of discrete actions is hand-tuned, particularly in domains where the actions are complex. Trainee Mansley was able to share this and related approaches to reinforcement learning in the classroom in a new course that teaches the mathematical underpinnings of computer science to college freshmen who are not majors math or science.

- “Research Achievements”
- Achievements for this Project