অরিন্দম ঘোষ

Research Interests

• Artificial Intelligence
• Machine Learning
• Deep Reinforcement Learning
• Multi-Agent Systems
• Evolutionary Algorithms
• Swarm Robotics
• Swarm Intelligence

Current Research / Ongoing Projects

• Graph Attention-based Actor-Critic Framework in Multi-Agent Systems for Role Adaptation and Improved Area Coverage
  My doctoral research focuses on developing scalable, adaptive, and decentralized decision making for coordinating multiple autonomous agents in disaster environments. It uses multiple critics to train the agent for wide-area coverage role adaptation, thereby improving mission efficiency in simulated SAR environments.
• Hybrid GA-DRL Framework for Area Coverage in uncertain environment
  This project investigates a hybrid framework for wide-area search and rescue that integrates a single genetic algorithm initialization with a deep reinforcement learning controller. This approach maximizes area coverage, minimizes overlap, improves victim detection, and adapts as team size and environmental complexity increase.
• LLM-Integrated Maze Problem Solving
  This project investigates how Large Language Models can be integrated into a reinforcement learning framework to assist agents in solving maze environments. The objective is to combine spatial reasoning and visual pathfinding to improve generalization across unseen maze topologies.
• Flocking-Based Experiment for Decentralized Training
  Inspired by natural swarms, this research explores how agents can learn coordinated behaviors like flocking using decentralized training paradigms. It emphasizes independent learning without centralized control, aiming to understand the scalability and robustness of such approaches in large-agent systems.

Past Research / Mini-Projects

• Adaptive Adversarial Learning in Multi-Agent Combat
  To address adversarial overfitting in multi-agent combat, this project developed a curriculum-based Deep Reinforcement Learning framework that employs deception strategies against evolving enemy evasion tactics. By integrating Deep Q-Networks (DQN) with adaptive curriculum learning in a PettingZoo environment, the system enables autonomous aircraft to generalize pursuit policies across diverse adversarial maneuvers.
• Couzin's Model for Flocking Behavior
  Implemented and analyzed the Couzin model to simulate emergent flocking behavior in autonomous agents. The project studied how local interaction rules among agents give rise to collective motion patterns in dynamic environments.
• Collective Decision-Making in Swarms
  Explored decentralized decision-making mechanisms in swarm systems using models like the majority rule and voter model. The study evaluated how consensus is achieved in noisy and uncertain environments through distributed interactions.
• Multi-Robot Path Planning for Area Coverage in Complex Environments
  Proposed an improved approach for multi-robot coverage path planning (mCPP) using A* and spanning tree coverage algorithms. The method efficiently partitions the workspace among robots and plans optimal single-robot paths to ensure complete coverage, even in environments with obstacles.
• Comparative Study on Grid-Based and Non-Grid-Based Path Planning Algorithms
  Conducted a comparative analysis of sampling-based path planning techniques including A*, PRM and RRT, with a focus on the effectiveness of grid-based sampling in reducing computational complexity while maintaining exploration efficiency.
• Modeling and Prediction of COVID-19 in India Using Machine Learning
  Developed a forecasting framework using non-linear regression, recurrent neural networks, and artificial neural networks to predict COVID-19 spread in India. The proposed model demonstrated improved long-term prediction accuracy and offered insights into the impact of public health interventions.

Collaborations & Mentoring

I actively collaborate with researchers in AI, robotics, and optimization. I also mentor undergraduate and graduate students on research projects in swarm intelligence and reinforcement learning.