rl competition

This is my solution for RLCOMP2020 reaching top 10. We used many approaches and different techniques including variations of deep q learning, imitation learning. After a month of testing and researching, I found that the key component of initial agent is the use of CNN. Because of difference between maps, so using 2D state would result a robust performance and generalize well across maps. Secondly, I tested many smaller combinations of RAINBOW algorithm, and the conbination including double deep q learning, prioritized experience memory, n-step learning give the most promising metrics. I aslo wrote a code that implement the learning from human demonstation paper. My teamates coded a heuristic bot based on A* algorithm then I used it as a demonstration generator instead of a human. However, due to lack of memory so this approach hasn't evalutated but I considered it as very promising approach.