Deep Reinforcement Learning algorithm to defeat a human player in Air Hockey Game. This is the capstone project of Hyojeong Kim, Namki Yu and Kyungeun Kim in Hanyang University.
Using Deep Q Network
-
hyper parameters
- discount_factor = 0.99
- learning_rate = 0.001
- epsilon = 1.0
- epsilon_decay_step = 9e-06
- batch_size = 64
- train_start = 1000
-
neural network
- 4 Layers
- input size : 8
- paddle_position
- paddle_velocity
- puck_location
- puck_velocity
- output size : 9
- up, down, right, left, up_left, down_left, up_right, down_right, stop
- hidden size : (24, 24)
- initializer : 'he_uniform'
- loss : 'mse'
- optimizer : 'Adam'
Goal : Using OpenAI Gym Environment
- Add goal area
- Add 'stop' action
- Change CNN to DNN