Research Areas

Human-Guided Reinforcement Learning

Developing algorithms that enable humans to effectively train and shape autonomous agents through demonstrations, real-time interventions, evaluative feedback, and natural language guidance. This includes foundational work on interactive agent shaping (Deep TAMER), the Cycle-of-Learning framework, and rating-based reinforcement learning.

Deep Reinforcement Learning & Multi-Agent Systems

Exploring deep reinforcement learning techniques for complex environments including video games (Atari, StarCraft II, Minecraft) and multi-agent coordination for human-robot collaboration and autonomous systems in tactical environments.

Large Language Models & Generative AI

Applying human-guided machine learning principles to LLMs for command and control planning, leveraging human feedback to steer and align generative AI systems.

Featured Projects

Deep TAMER

Extended interactive agent shaping to high-dimensional state spaces using deep learning. Enables non-expert humans to train agents in complex environments like Atari games by providing evaluative feedback (positive/negative signals) rather than explicit demonstrations. Published at AAAI 2018.

Human-in-the-Loop RL Deep Learning AAAI

Publication

Cycle-of-Learning

A unified framework integrating multiple human interaction modalities—demonstrations, real-time interventions, and evaluative feedback—into reinforcement learning. Defines switching criteria between modalities for efficient autonomous system training. Published at AAAI 2019.

Human-Guided RL Imitation Learning AAAI

arXiv AAAI

Narration-Guided RL for StarCraft II

A novel approach using natural language narration to address reward sparsity in deep RL. Projects language commands into a shared representation with goal states, enabling agents to learn tasks that were previously unlearnable in complex environments like StarCraft II.

Language-Guided RL StarCraft II NLP

arXiv

Rating-based Reinforcement Learning

A framework allowing humans to train agents using intuitive numerical ratings rather than binary feedback or demonstrations. Extends the types of human feedback that can be leveraged for RL. Published at AAAI 2024.

Human Feedback Reinforcement Learning AAAI

Publication

GUIDE: Real-Time Human-Shaped Agents

Enables continuous human feedback grounded into dense rewards. Features a simulated feedback module that learns to replicate human guidance patterns. With only 10 minutes of human feedback, achieves up to 30% increase in success rate. Published at NeurIPS 2024.

Real-Time RL Human Feedback NeurIPS

Publication

Multi-Agent Curriculum Learning in StarCraft II

Learning to guide multiple heterogeneous actors from a single human demonstration via automatic curriculum learning. Addresses the challenge of multi-agent coordination from limited human input.

Multi-Agent RL Curriculum Learning StarCraft II

Publication

DIP-RL: Demonstration-Inferred Preference Learning

Novel approach to learning from human demonstrations in Minecraft environments. Bridges the gap between human intuition and AI learning through preference inference.

Preference Learning Human Feedback Minecraft

arXiv

COA-GPT

Applying human-guided ML principles to LLMs for accelerated Course of Action development in military operations. Demonstrates how RL concepts extend to generative AI for Command and Control planning.

LLMs Generative AI C2 Planning

Publication

Collaborations & Competitions

• MineRL BASALT 2022 Competition - Contributed to solving fuzzy tasks with human feedback in Minecraft environments
• NeurIPS 2021 Competition - Retrospective on learning from human feedback
• AAAI, ICML, NeurIPS, AAMAS - Regular presenter and contributor to top-tier AI/ML conferences

---

← Back to Home