ray.init()
results = {}
N = 100
config["num_workers"] = 1
config["num_gpus"] = 0
# You may have to run each agents in separate sessions
# to avoid PyBullet restrictions
agent = "ALP-GMM"
# agent = "Manual"
# agent = "No Curriculum"
print(f"Evaluating agent: {agent}")
results[agent] = []
trainer = PPOTrainer(config=config, env=envs[agent])
trainer.restore(agents[agent])
env = envs[agent](dict(config["env_config"], **{"in_training": False}))
for i in range(N):
print(agent, i)
done = False
obs = env.reset()
ep_reward = 0
while not done:
action = trainer.compute_action(obs)
obs, reward, done, info = env.step(action)
ep_reward += reward
if done:
obs = env.reset()
results[agent].append(ep_reward)
print(f"Agent {agent} score: {np.round(np.mean(results[agent]), 2)}")
trainer.restore(
# Replace this with your checkpoint path.
"/home/enes/ray_results/PPO_InventoryEnv_2020-10-06_04-58-04t8r36o9o/checkpoint_781/checkpoint-781"
)
if __name__ == "__main__":
np.random.seed(0)
env = InventoryEnv()
episode_reward_avgs = []
episode_total_rewards = []
for i in range(2000):
print(f"Episode: {i+1}")
state = env.reset()
done = False
ep_rewards = []
while not done:
action = trainer.compute_action(state)
state, reward, done, info = env.step(action)
ep_rewards.append(reward)
total_reward = np.sum(ep_rewards)
reward_per_day = np.mean(ep_rewards)
print(f"Total reward: {total_reward}")
print(f"Reward per time step: {reward_per_day}")
episode_reward_avgs.append(reward_per_day)
episode_total_rewards.append(total_reward)
print(
f"Average daily reward over {len(episode_reward_avgs)} "
f"test episodes: {np.mean(episode_reward_avgs)}. "
f"Average total epsisode reward: {np.mean(episode_total_rewards)}"
)
ray.init()
num_policies = 4
policies = {
"policy_{}".format(i): (None, env.observation_space, env.action_space, {})
for i in range(num_policies)
}
policy_ids = list(policies.keys())
config = {
"multiagent": {
"policies": policies,
"policy_mapping_fn": (lambda agent_id: random.choice(policy_ids)),
},
"framework": "tf",
}
#trainer = ApexTrainer(env=TicTacToe, config=config)
trainer = PPOTrainer(env=TicTacToe, config=config)
trainer.restore("ttt_model/checkpoint_51/checkpoint-51")
obs = env.reset()
print(obs)
done = False
while not done:
env.render()
player = list(obs)[0]
if player == "X":
action = int(input(f"Player {player} - enter action 1-9:")) - 1
else:
action = trainer.compute_action(np.array(obs["O"]), policy_id="policy_1")
obs, rewards, dones, infos = env.step({player: action})
done = dones["__all__"]
print(obs, rewards, dones, infos)
env.render()
steps = 0
draw = 0
wins = 0
losses = 0
# Select no. of episodes
n_episodes = 100
avg_rew = 0
total_rew = 0
# Play game with trained agent
for i_episode in range(1, n_episodes + 1):
obs = env.reset()
episode_rew = 0
while True:
# Sample action from trained policy
action = ppo_trainer.compute_action(obs, policy_id='policy_01')
action_dict[str(int(action))] += 1
obs, rew, done, info = env.step(action)
episode_rew += rew
if done:
break
if episode_rew >= 0 and episode_rew <= 1.0:
draw += 1
if episode_rew > 1.0:
wins += 1
total_rew += episode_rew
print("Episode:%d Reward:%.2f" % (i_episode, episode_rew))
avg_rew = total_rew / i_episode
