done = False
rewards_current_episode = 0
print(f"{episode + 1} / {num_episodes}")
# Play episode
for step in range(max_steps_per_episode):
# Exploration vs Exploitation
exploration_rate_threshold = random.uniform(0, 1)
if exploration_rate_threshold > exploration_rate:
action = np.argmax(q_table[state, :])
else:
action = np.random.choice([x for x in range(action_space_size)])
# Step
new_state, reward, done, feedback = env.step(action)
new_state = state_to_coord(new_state)
# Update Q table
q_table[state, action] = q_table[state, action] * (1 - learning_rate) + \
learning_rate * (reward + feedback + discount_rate * np.max(q_table[new_state, :]))
# state = new state and accumulate rewards
state = new_state
rewards_current_episode += reward
if done == True:
break
# End of episode
# Update exploration rate
exploration_rate = min_exploration_rate + \
# Pre-train
#agent.replay.tree.start = start
for i in range(k1):
if i % 100 == 0:
print("pretraining:", i)
agent.learn()
# Train
accumulated_rewards_all_episodes = []
for episode in range(k2):
s = env.reset()
accumulated_rewards = 0
done = False
while not done:
a = agent.choose_action(s)
s_, r, done, feedback = env.step(a[0])
accumulated_rewards += r
r += feedback
if done:
r = 0 # todo not sure if this is necessary, just try with, without, with different value
agent.store_transition(s, a, r, s_, done)
agent.learn()
s = s_
# Update exploration rate
"""
agent.eps = min_exploration_rate + \
(max_exploration_rate - min_exploration_rate) * np.exp(-exploration_decay_rate * episode)