Research

LLM-Augmented Reinforcement Learning for Dynamic Decision-Making

Institution: Clemson University — Clemson, SC

Dates: Aug. 2024 - Present

• Developed a modular framework integrating domain-specific and general-purpose LLMs to support RL agents through context-aware prompting.
• Designed a dual-prompt learning mechanism that combines static domain-informed inputs with trainable prompt tokens for adaptable decision-making in dynamic environments.
• Implemented transformer-based models, including BERT, to extract contextual embeddings for adaptive prompt tuning and decision-making.
• Investigating improvements in sample efficiency and policy generalization across multi-agent RL tasks.
• The framework enables flexible multi-agent interaction and supports transferability across domains in complex decision-making systems.

Meta-Learning for Generalized Policy Optimization

Institution: Clemson University — Clemson, SC

Dates: Aug. 2023 - Present

• Designed a Model-Agnostic Meta-Learning (MAML) framework for improving generalization across diverse tasks.
• Applied few-shot learning principles using PyTorch Lightning to boost cross-task adaptability in non-stationary environments.
• Utilized Pandas and PySpark for preprocessing interaction logs and training data streams.

Multi-Agent Deep Reinforcement Learning for Coordinated Decision-Making

Institution: Clemson University — Clemson, SC

Dates: Aug. 2022 - Present

• Designed a sharpness-aware optimization strategy to stabilize multi-agent DRL training and enhance policy robustness.
• Demonstrated significant improvements in convergence speed and generalization across complex multi-task scenarios.
• Results under review at IEEE Transactions on Machine Learning in Communications and Networking as “Sharpness-Aware O-RAN Resource Management Using Multi-Agent Reinforcement Learning.”

Deep Metric Learning for Robust Feature Extraction

Institution: Clemson University — Clemson, SC

Dates: Aug. 2022 - May. 2023

• Engineered a contrastive loss framework for learning high-quality embeddings in low-data regimes.
• Applied this to imbalanced classification problems and validated improvements in class-wise F1-score and embedding separability.

Inverse Reinforcement Learning for Reward Modeling

Institution: Clemson University — Clemson, SC

Dates: Aug. 2022 - Aug. 2023

• Trained an IRL agent using expert trajectory data to infer optimal reward functions for unknown environments.
• Applied to sequential decision-making with real-world-inspired behavior policies.

Attention-Based Multi-Task Learning with Transformer Architectures

Institution: Clemson University — Clemson, SC

Dates: Jan. 2023 - Dec. 2023

• Applied attention mechanisms for adaptive loss weighting in RL frameworks, enabling dynamic prioritization across tasks in multi-agent settings, and leveraging shared transformer-based encoders for enhanced coordination.
• Achieved harmonized task optimization and reduced dominant modality bias.

Time-Series Forecasting using LSTM Networks

Institution: Clemson University — Clemson, SC

Dates: Aug. 2022 - Aug. 2023

• Implemented LSTM and GRU-based architectures to model user behavior trends and traffic dynamics.
• Preprocessed real-world sequence data using pandas and PySpark, optimizing prediction accuracy for long-tail temporal dependencies.

Semantic-Aware Reinforcement Learning for Contextual Decision Making

Institution: University of Colorado Colorado Springs — Colorado Springs, CO

Dates: Jan. 2021 - Aug. 2022

• Developed context-aware RL models that improve decision efficiency in dynamic environments through semantic feature extraction.