game.new_episode()
game_state = game.get_state()
misc = game_state.game_variables # [KILLCOUNT, AMMO, HEALTH]
prev_misc = misc
action_size = game.get_available_buttons_size()
img_rows, img_cols = 64, 64
img_channels = 3 # Color channel
trace_length = 4 # Temporal Dimension
state_size = (trace_length, img_rows, img_cols, img_channels)
agent = DoubleDQNAgent(state_size, action_size, trace_length)
agent.model = Networks.drqn(state_size, action_size, agent.learning_rate)
agent.target_model = Networks.drqn(state_size, action_size,
agent.learning_rate)
s_t = game_state.screen_buffer # 480 x 640
s_t = preprocessImg(s_t, size=(img_rows, img_cols))
is_terminated = game.is_episode_finished()
# Start training
epsilon = agent.initial_epsilon
GAME = 0
t = 0
max_life = 0 # Maximum episode life (Proxy for agent performance)
life = 0
episode_buf = [] # Save entire episode
