Quantum Mean-field Multi Agent Reinforcement Learning
By Rishi
My project aims to create a learning framework for large multi-agent systems that combines mean field game theory with model-free quantum reinforcement learning methods. This means formulating a representative agent model, where the agent would interact with the rest of the population, and their interaction would be described by the reduced quantum states or the expectation-level dynamics, which would then allow for the use of the quantum abstractions such as the quantum channels and the DMFT reduction. In addition, the use of the two-timescale update method would allow for both policy learning and the agent-level mean-field updates. This will also involve the usage of stabilizing components such as smooth stochastic policy, replay buffers, and target networks, coming from the classical and the quantum reinforcement learning. This would help develop an approximate model for the mean field equilibrium without needing any information regarding the external environment, providing an efficient method for learning control and simulations involving the use of the hybrid quantum-classical multi-agent populations. This would hopefully be applicable for the wide range of the quantum decision-making processes, specifically for finance and the economics, as well as the new quantum-enabled simulations involving the use of the many agents.