Kai Widdeson

Gravitar Coursework

This coursework required us to use Reinforcement learning to create an agent capable of playing the arcade game Gravitar. Gravitar is notoriously difficult for agents to perform well on. My final implementation involved a duelling DQN with noisy layers, loss of life penalty, and prioritised experience replay with a target network. Experimentaton was carried out with a Double Duelling DQN and Multistep learning however these had negative impacts on performance. A video of the best runthrough and a graph showing the agent's learning process can be found below.