Modelling Adversary Intent Using Multiobjective Reinforcement Learning
This project will develop a framework that uses multi-player adversarial games to demonstrate the capacity of an agent to achieve its own goals within an uncertain environment while adapting those plans based on evolving beliefs about the intentions of its adversaries. The games chosen will allow each player to select between various goals which all offer a pathway towards achieving the overall intent of the player (to achieve one of a number of available win conditions). This framework allows for modelling of the intended goals of an opponent, and the exploitation of this information to better achieve the agent’s own objectives. By also incorporating elements of uncertainty into the game (such as hidden state, asymmetries in the availability of state information, stochastic outcomes of actions, and partial observability of opponent actions), this framework will also be suitable for demonstrating the capacity to exploit opponent modelling even when there is significant uncertainty within the environment.
Funding Agency/Company: DSTG $53,000