Reinforcement Learning Game Player
-
The project involved building a reinforcement learning agent to effectively solve the LunarLander v2.0 environment on OpenAI Gym.
-
The environment involved a floating platform which changed position from episode to episode, a booster rocket and the lunar surface. The objective was to successfully land the booster rocket on the floating platform. The actions available to the agent were UP, DOWN, LEFT, and RIGHT.
-
The primary challenge in designing this project was to decide the kind of agent to be used. I utilized a double deep Q-network as opposed to a deep Q-network for mitigating any possible information bleed back to the neural network.
-
The double deep network essentially utilizes a network for training to make the decisions, and then transfers the learning to a target network at test, thus mitigating possibilities of information bleed.
-
The entire projecct was written up in Python with the Keras library and a Theano back end.
