plt.figure(figsize=[10, 4])
rewards = []
# Training loop
for i in range(max_iterations):
# Play & train game
# Update rewards
# rewards
# Decay agent epsilon
# agent.epsilon = ?
initial_state = env.reset()
rewards.append(play_and_train(env, agent))
agent.epsilon = agent.epsilon * discount
if i % 100 == 0:
print('Iteration {}, Average reward {:.2f}, Epsilon {:.3f}'.format(
i, np.mean(rewards), agent.epsilon))
if visualize:
plt.subplot(1, 2, 1)
plt.plot(rewards, color='r')
plt.xlabel('Iterations')
plt.ylabel('Total Reward')
plt.subplot(1, 2, 2)
plt.hist(rewards,
bins=20,
range=[-700, +20],
discount = 0.99
agent = QLearningAgent(alpha, epsilon, discount, getActionRange)
agent2 = SarsaAgent(alpha, epsilon, discount, getActionRange)
plt.figure(figsize=[10, 4])
rewards1 = []
rewards2 = []
# Training loop
for i in range(max_iterations):
# Play & train game
rewards1.append(play_and_train(env, agent))
rewards2.append(play_and_train(env, agent2))
if (i + 1) % 100 == 0:
agent.epsilon = max(agent.epsilon * 0.99, 0.00001)
agent2.epsilon = max(agent2.epsilon * 0.99, 0.00001)
# agent.alpha = max(agent.alpha * 0.99, 0.00001)
# agent2.alpha = max(agent2.alpha * 0.99, 0.00001)
if i % 100 == 0:
print(
'Iteration {}, Average reward {:.2f}, Average reward {:.2f}, Epsilon {:.3f}'
.format(i, np.mean(rewards1[-100:]), np.mean(rewards2[-100:]),
agent.epsilon))
if visualize:
plt.subplot(1, 2, 1)
plt.plot(rewards1, color='r')
plt.plot(rewards2, color='b')
plt.xlabel('Iterations')