Scrappy/Simpkins Lab for Interactive AI

CCSE Research Showcase 2026-01-15

Vision: Competent, Flexible, Human-Aligned Intelligent Autonomous Agents

• Multiagent reinforcement learning
• Multiple goal reinforcement learning
• Model-based reinforcement learning

"Thinking is nothing but acting in an imagined space."

-- Konrad Lorenz via Bernhard Schölkopf

• Offline reinforcement learning
• Top-to-bottom layered autonomous intelligent agent architecture
  • Perception
  • Reasoning
  • Decision making
  • Action

Imagine an autonomous intelligent agent that learns "how to behave" by watching thousands of hours of movies.

Ends don't justify means: learning action restrictions from a teacher

• Agent in red demonstrates "optimal" path to goal.
• Student agent is blue.
• Shaded area is "restricted."
• Student learned from observations of a teacher who avoided restricted area.
• Student avoids restricted area, even if it results in suboptimal goal attainment.

Evaluating Advancements in Reinforcement Learning in Video Games

Untrained boxer agent	Trained boxer agent