ball_y = state[1]
ball_speed_x = state[2]
ball_speed_y = state[3]
paddle_x = state[4]
index += 2000 * ball_speed_x
index += 1000 * ball_speed_y
index += 100 * ball_x
index += 10 * ball_y
index += 1 * paddle_x
return index
# Init environment
env = Game()
# Create Q table
action_space_size = env.n_actions
state_space_size = env.n_states_disc
q_table = np.zeros(
(state_space_size, action_space_size)) # n_states x n_actions
# Hyper parameters
num_episodes = 10000
max_steps_per_episode = 1000000
learning_rate = 0.1 # alpha
discount_rate = 0.99 # gamma
exploration_rate = 1 # epsilon
lambda_2 = 1.0
lambda_3 = 1e-5
l_margin = 0.8
e_d = 0.005 # bonus for demonstration # todo try with 1 as that is what the paper said
e_a = 0.001
# Environment Parameters
lr = 0.002
batch_size = 64
epsilon = 1.0 # Decreased over time
max_exploration_rate = 1
min_exploration_rate = 0.001
exploration_decay_rate = 0.01
# Initialize Environment
env = Game()
# Initialize Agent
agent = Agent(lr=lr,
eps=epsilon,
gamma=gamma,
max_memory=max_memory,
n_steps=n_steps,
batch_size=batch_size,
tau=tau,
lambda_1=lambda_1,
lambda_2=lambda_2,
lambda_3=lambda_3,
l_margin=l_margin)
# Load Demonstration Data