num_states = environment.observation_space.n
num_actions = environment.action_space.n
agent = QAgent(num_states, num_actions)
sum_reward = 0
for episode in range(NUM_EPISODES):
done = False
last_state = environment.reset()
last_reward = None
# Number of steps taken. A bit of a safeguard...
num_steps = 0
while not done:
# Epsilon-greedy policy
action = agent.get_action(last_state, environment)
state, reward, done, info = environment.step(action)
# A crude timeout: If we play too long without
# completing the level, kill the game
num_steps += 1
if num_steps > 1000:
print(
"Episode timeout! Could not finish in 1000 steps. Check your actions!"
)
done = True
# Update Q-table if we have one whole experience of
# s, a, r, s', t'
if last_state is not None:
num_states = environment.observation_space.n
num_actions = environment.action_space.n
vtable = VTable(num_states, discount_factor=0.5)
agent = QAgent(num_states, num_actions)
# Load already trained Q-table
agent.load("q_table.npy")
for episode in range(NUM_EPISODES):
done = False
state = environment.reset()
# Keep track of visited states and rewards
# obtained
states = []
rewards = []
while not done:
# Store state
states.append(state)
# Take action according to Q-agent
action = agent.get_action(state, environment)
state, reward, done, info = environment.step(action)
# Store reward
rewards.append(reward)
# Update v-estimate with the played game
vtable.process_trajectory(states, rewards)
if ((episode + 1) % SHOW_EVERY_EPISODES) == 0:
vtable.visualize_v((4, 4))