def train(env, env_name, agent, episodes, rpm):
max_reward = -1e10
while len(rpm) < opt["MEMORY_WARMUP_SIZE"]:
run_episode(env, agent, rpm)
for i in range(episodes):
step, total_reward = run_episode(env, agent, rpm)
writer.add_scalar(env_name, total_reward, i)
if i % 10 == 0:
print("Episode {}, step {} Reward Sum {}.".format(i, step, total_reward))
logging.warning("Episode {}, step {} Reward Sum {}.".format(i, step, total_reward))
if (i + 1) % 100 == 0:
total_reward = evaluate(10, env, agent, render=False)
if total_reward >= max_reward:
max_reward = total_reward
agent.save(env_name)
def do_save(self, args):
agent.save()
# average angle in episode
scores.append(score)
if len(scores) > 20: scores = scores[-20:]
if len(rewards) > 20: rewards = rewards[-20:]
if highest_score == None or score < highest_score:
highest_score = score
print(
"episode: {}/{}, average angle: {}, best average angle: {}, last 20 average: {}, e: {}, in memory: {}, batch size: {}"
.format(e + 1, episode_count, round(score, 2),
round(highest_score, 2), round(mean(scores), 2),
round(agent.epsilon, 3), len(agent.memory), batch_size))
if len(scores) >= 5 and score <= 10 and sum(scores[-5:]) <= 50:
agent.save()
if len(scores) >= 15 and sum(scores[-15:]) <= 75:
print("training successfull!")
agent.save("final")
break
if (e + 1) % 5 == 0:
print("Took", round((time.time() - start) / 60, 2), "minutes\n")
start = time.time()
agent.merge_models()
agent.save()
print("Total training time:", round((time.time() - first_start) / 60, 2),
"minutes")
env.print_info()
#agent = agent.Agent(env)
agent = agent_dqn.DQNAgent(env, "arkanoid_network.json")
training_iterations = 250000
for i in range(0, training_iterations):
agent.main()
if (i % 100) == 0:
progress = 100.0 * i / training_iterations
print("training done = ", progress, " score = ", env.get_score())
agent.save("arkanoid_network/")
env.reset_score()
agent.run_best_enable()
testing_iterations = 10000
for i in range(0, testing_iterations):
agent.main()
print("********************************")
print("testing score = ", env.get_score())
print("********************************")
while True:
agent.main()
env.render()