Data Driven Policy Learning in Real World Multi-Agent Environments

The following is the tentative outline of the proposed tutorial. I will start with a brief overview of key concepts from Reinforcement learning and then extend the ideas to Multi-Agent settings. Along the way I will explain the key challenges for learning algorithm development in this area, and finally illustrate the application of the algorithms in practical applications.

1. Introduction to Data driven policy learning (20 Mins)

• Reinforcement learning – theoretical underpinnings

• Generalized policy iteration

• Deep reinforcement learning

• Inverse Reinforcement Learning and Imitation learning

• Applications of data driven policy learning

2. Applications of Policy learning in multi-agent environments (10 Mins)

• Requirement for multi-agent policy learning: through a discussion of applications

• Introduction to Multi-agent Reinforcement learning

• Theoretical foundations of Multi-agent learning

3. Open Challenges and Recent Developments in multi-agent policy learning (40 Mins)

• Curse of dimensionality and non-stationarity in multi-agent policy learning

• Extending Actor-Critic methods from single agent to multi-agent reinforcement learning

• Communication between agents and associated methods to overcome

• Imitation learning in multi-agent systems: Generative adversarial and coordinated methods

• Curriculum learning for efficient multi-agent learning

• Selective attention methods to overcome curse of dimensionality

4. Applications of data-driven multi-agent policy learning (20 Mins)

• Deep imitation learning for generating control policies in mixed autonomous-human driven traffic environments

• Predicting the team level movements in team sports such as football and basketball, and its use in performance monitoring

• Real world reinforcement learning of an autonomous vehicle, and testing driverless cars through multi-agent simulation

• Emerging applications of multi-agent policy learning for climate change prediction, wild life conservation, and financial technologies

• Experimental platforms for multi-agent policy learning