In multi agent environments, multiple agents perform actions in a joint environment, where the action of each agent affects other agents as well. In this project we developed a custom gym multi agent environment, where the agents are divided into two groups – attackers and defenders with opposite goals. We used two popular state of the art RL algorithms – PPO and IMPALA in order to train the attackers to inflict maximum damage on the defenders. We see that while both algorithms receive a similar mean reward after finishing the training, PPO has converged faster showing it has better sample complexity. Moreover, the minimal reward of PPO after running it on 100 randomly generated scenarios is much higher, showing that the trained policy is more robust than IMPALA.