Simultaneous Multiagent learning in an adversarial setting

This is the index of my BTP(B.Tech. Project) submissions. We picked up a game called 'Keepout' and implemented Reinforcement Learning algorithms for the agent to learn from their experience and improve over trials. We conducted various experiments and the results can be seen in the report.

Problem Statement

The game of Keepout consists of two teams - Attackers & Defenders - with at least one member per team present on the arena. Any one of the attackers have to reach a ball kept inside a target zone on the arena, and the defenders have to prevent the attackers from doing so, for as long as possible. In this process, the defender can block/impede any attacker's path, but not essentially push it in doing so. Similarly, the attacker may not push the defender. Each such run is limited by a timeframe T0, inside which each robot has to accomplish its task. Hence, each run lasts for at most T0 time, after which the run is aborted (in case any attacker is not able to reach the ball in that time), and a new run commences. For the purpose of simulation, it shall be assumed that any collision between attacker and defender while the trial is still running will mean defender has been able to successfully 'impede' the attacker and the run stops there.

proposal
BTP sem 1 report draft
BTP sem 2 report
BTP sem 2 presentation link
The code implementation in VC++ can be downloaded using this link