def exercise_agent(gymName, episodes, render=True, convolutional=False):
max_t = 0
env = gym.make(gymName)
agent = QAgent(env.action_space.n, convolutional)
for i_episode in range(episodes):
state = normalize(env.reset())
agent.observe(state, 0, False)
total_reward = 0
for t in range(10000):
if render:
env.render()
action = agent.act()
state, reward, done, info = env.step(action)
state = normalize(state)
total_reward += reward
agent.observe(state, reward, done)
if done:
max_t = max(max_t, t)
print(f'{t} : {max_t} : {total_reward}')
break
env.close()
env = BananaEnv()
agent = QAgent(action_space=env.get_action_space_size(),
state_space=env.get_state_space_size())
if load_bad_network:
agent.load_checkpoint(bad_opt_networks_ckp)
elif load_from_seeded_64: #target and delayer weights are not actually needed here, they would be needed to resume training as it was left.
agent.load_checkpoint(local_checkpoint=seeded_test_64_ckp["local"],
target_checkpoint=seeded_test_64_ckp["target"],
delayer_checkpoint=seeded_test_64_ckp["delayer"])
else:
agent.load_checkpoint(top_opt_networks_ckp[sel_network])
env.reset()
done = False
for i in range(5):
print("Episode {:d}\n score: ".format(i), end=" ")
done = False
env.reset()
score = 0
while not done:
exp = agent.act(env, 0)
score = score + exp.reward
if abs(exp.reward) > 0.01:
print(str(int(score)), sep=' ', end=' ', flush=True)
done = exp.done
sleep(0.02)
print("\nfinal score:" + str(score) + "\n")
sleep(1)
score_list = []
mean_score_list = []
running_score = 0
eps_start = 1.0
eps_decay = pars["eps_decay_sel"]
eps_end = 0.01
eps = eps_start
max_ep_len = 400
train_episodes = 701
for episode in range(train_episodes):
eps = max(eps * (eps_decay), eps_end)
env.reset()
done = False
curr_score = 0
for act_i in range(max_ep_len):
exp = agent.act(env, eps)
curr_score = curr_score + exp.reward
if exp.done:
break
if act_i % learn_every == 0:
agent.learn(64)
score_list.append(curr_score)
score_window.append(curr_score)
if episode % update_every == 0:
agent.update_target()
if episode % 20 == 0:
print("episode " + str(episode) + ", mean score: " +
str(np.mean(score_window)))
if episode % 100 == 0:
mean_score_list.append(np.mean(score_window))
print("test completed with scores: " + str(mean_score_list))
import gym
from q_agent import QAgent
env = gym.make('FrozenLake-v0')
print(env.action_space)
print(env.observation_space)
agent = QAgent(env.observation_space, env.action_space);
agent.learn(env)
success = 0
for i_episode in range(100):
observation = env.reset()
while True:
#env.render()
action = agent.act(observation)
observation, reward, done, info = env.step(action)
if done:
#print("Episode finished after {} timesteps".format(t+1))
if reward == 1.0:
success += 1
break
print("success rate is {}".format(success))