def test_vpg_agent(self):
config = {
'batch_size': 8,
'max_episode_length': 4,
'continuous': False,
'state_shape': (2,),
'actions': 2}
tf.reset_default_graph()
config = create_config(config)
network_builder = NeuralNetwork.layered_network(layers=[{'type': 'dense', 'num_outputs': 32}])
agent = VPGAgent(config=config, network_builder=network_builder)
state = (1, 0)
rewards = [0.0] * 100
for n in range(10000):
action = agent.get_action(state=state)
if action == 0:
state = (1, 0)
reward = 0.0
terminal = False
else:
state = (0, 1)
reward = 1.0
terminal = True
agent.add_observation(state=state, action=action, reward=reward, terminal=terminal)
rewards[n % 100] = reward
if sum(rewards) == 100.0:
return
self.assertTrue(False)
def test_baseline(self):
config = Configuration(discount=0.75,
batch_size=8,
learning_rate=0.001,
states=dict(shape=(1, )),
actions=dict(continuous=True),
network=layered_network_builder(()))
agent = VPGAgent(config=config)
states = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
rewards = [0.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 1.0, 0.0]
terminals = [
False, False, False, False, True, False, False, False, True
]
discounted_rewards = np.array([
0.75 + 0.75**4, 1.0 + 0.75**3, 0.75**2, 0.75, 1.0, 1.0 + 0.75**2,
0.75, 1.0, 0.0
])
baseline = np.array([0.25, 0.5, 0.0, 0.25, 0.5, 0.5, 0.25, 0.5, 0.0])
agent.model.baseline = dict(state=Baseline())
agent.model.baseline['state'].predict = lambda states: baseline
result, _ = agent.model.reward_estimation(states=dict(state=states),
rewards=rewards,
terminals=terminals)
expected = discounted_rewards - baseline
print(result)
print(expected)
self.assertTrue((result == expected).all())
def test_multithreaded(self):
sys.stdout.write('\nVPGAgent (multithreaded):')
sys.stdout.flush()
environment = MinimalTest(specification={'int': ()})
network = [dict(type='dense', size=32), dict(type='dense', size=32)]
kwargs = dict(update_mode=dict(unit='episodes',
batch_size=4,
frequency=4),
memory=dict(type='latest',
include_next_states=False,
capacity=100),
optimizer=dict(type='adam', learning_rate=1e-2))
agent = VPGAgent(states=environment.states,
actions=environment.actions,
network=network,
**kwargs)
agents = clone_worker_agent(agent, 5, environment, network, kwargs)
environments = [environment
] + [copy.deepcopy(environment) for n in range(4)]
runner = ThreadedRunner(agent=agents, environment=environments)
runner.run(num_episodes=100)
runner.close()
sys.stdout.write(' ran\n')
sys.stdout.flush()
def test_beta(self):
passed = 0
for _ in xrange(5):
environment = MinimalTest(definition=True)
actions = environment.actions
actions['min_value'] = -0.5
actions['max_value'] = 1.5
config = Configuration(batch_size=8,
learning_rate=0.01,
states=environment.states,
actions=actions,
network=layered_network_builder([
dict(type='dense', size=32),
dict(type='dense', size=32)
]))
agent = VPGAgent(config=config)
runner = Runner(agent=agent, environment=environment)
def episode_finished(r):
return r.episode < 100 or not all(x / l >= 0.9 for x, l in zip(
r.episode_rewards[-100:], r.episode_lengths[-100:]))
runner.run(episodes=1500, episode_finished=episode_finished)
print('VPG agent (beta): ' + str(runner.episode))
if runner.episode < 1500:
passed += 1
print('VPG agent (beta) passed = {}'.format(passed))
self.assertTrue(passed >= 4)
def test_continuous(self):
passed = 0
for _ in xrange(5):
environment = MinimalTest(definition=True)
config = Configuration(batch_size=8,
learning_rate=0.001,
states=environment.states,
actions=environment.actions,
network=layered_network_builder([
dict(type='dense', size=32),
dict(type='dense', size=32)
]))
agent = VPGAgent(config=config)
runner = Runner(agent=agent, environment=environment)
def episode_finished(r):
return r.episode < 100 or not all(
x >= 1.0 for x in r.episode_rewards[-100:])
runner.run(episodes=1000, episode_finished=episode_finished)
print('VPG agent (continuous): ' + str(runner.episode))
if runner.episode < 1000:
passed += 1
print('VPG agent (continuous) passed = {}'.format(passed))
self.assertTrue(passed >= 4)
def test_discrete_baseline(self):
passed = 0
for _ in xrange(5):
environment = MinimalTest(definition=False)
config = Configuration(batch_size=8,
learning_rate=0.001,
states=environment.states,
actions=environment.actions,
baseline=dict(type="mlp",
sizes=[32, 32],
epochs=5,
update_batch_size=8,
learning_rate=0.01),
network=layered_network_builder([
dict(type='dense', size=32),
dict(type='dense', size=32)
]))
agent = VPGAgent(config=config)
runner = Runner(agent=agent, environment=environment)
def episode_finished(r):
return r.episode < 100 or not all(x / l >= 0.9 for x, l in zip(
r.episode_rewards[-100:], r.episode_lengths[-100:]))
runner.run(episodes=1500, episode_finished=episode_finished)
print('VPG agent (discrete): ' + str(runner.episode))
if runner.episode < 1500:
passed += 1
print('VPG agent (discrete) passed = {}'.format(passed))
self.assertTrue(passed >= 4)
def test_multi(self):
passed = 0
def network_builder(inputs):
layer = layers['dense']
state0 = layer(x=layer(x=inputs['state0'], size=32), size=32)
state1 = layer(x=layer(x=inputs['state1'], size=32), size=32)
state2 = layer(x=layer(x=inputs['state2'], size=32), size=32)
state3 = layer(x=layer(x=inputs['state3'], size=32), size=32)
return state0 * state1 * state2 * state3
for _ in xrange(5):
environment = MinimalTest(definition=[
False, (False, 2), (False, (1, 2)), (True, (1, 2))
])
config = Configuration(batch_size=8,
learning_rate=0.001,
states=environment.states,
actions=environment.actions,
network=network_builder)
agent = VPGAgent(config=config)
runner = Runner(agent=agent, environment=environment)
def episode_finished(r):
return r.episode < 50 or not all(
x >= 1.0 for x in r.episode_rewards[-50:])
runner.run(episodes=2000, episode_finished=episode_finished)
print('VPG agent (multi-state/action): ' + str(runner.episode))
if runner.episode < 2000:
passed += 1
print('VPG agent (multi-state/action) passed = {}'.format(passed))
self.assertTrue(passed >= 4)
def get_agent(agentType):
if agentType == "dqn":
agent = DQNAgent(
states={
"type": 'float',
"shape": (
int(args.population),
1,
int(args.resources),
)
},
actions={
"type": 'int',
"shape": (int(args.resources), ),
"num_values": 3
},
memory=1000,
network="auto",
)
elif agentType == "vpg":
agent = VPGAgent(
states={
"type": 'float',
"shape": (
int(args.population),
1,
int(args.resources),
)
},
actions={
"type": 'int',
"shape": (int(args.resources), ),
"num_values": 3
},
network="auto",
memory=1000,
)
elif agentType == "trpo":
agent = TRPOAgent(
states={
"type": 'float',
"shape": (
int(args.population),
1,
int(args.resources),
)
},
actions={
"type": 'int',
"shape": (int(args.resources), ),
"num_values": 3
},
network="auto",
memory=1000,
)
return agent
def test_multi_baseline(self):
passed = 0
def network_builder(inputs, **kwargs):
layer = layers['dense']
state0 = layer(x=layer(x=inputs['state0'],
size=32,
scope='state0-1'),
size=32,
scope='state0-2')
state1 = layer(x=layer(x=inputs['state1'],
size=32,
scope='state1-1'),
size=32,
scope='state1-2')
state2 = layer(x=layer(x=inputs['state2'],
size=32,
scope='state2-1'),
size=32,
scope='state2-2')
return state0 * state1 * state2
for _ in xrange(5):
environment = MinimalTest(
definition=[False, (False, 2), (True, 2)])
config = Configuration(batch_size=8,
learning_rate=0.001,
baseline=dict(type="mlp",
sizes=[32, 32],
epochs=5,
update_batch_size=8,
learning_rate=0.01),
states=environment.states,
actions=environment.actions,
network=network_builder)
agent = VPGAgent(config=config)
runner = Runner(agent=agent, environment=environment)
def episode_finished(r):
return r.episode < 100 or not all(
x / l >= reward_threshold for x, l in zip(
r.episode_rewards[-100:], r.episode_lengths[-100:]))
runner.run(episodes=4000, episode_finished=episode_finished)
print('VPG agent (multi-state/action): ' + str(runner.episode))
if runner.episode < 4000:
passed += 1
print('VPG agent (multi-state/action) passed = {}'.format(passed))
self.assertTrue(passed >= 4)
def __init__(self, state_size, env=None, is_eval=False):
self.state_size = state_size
self.action_size = 3
self.memory_size = 1000
self._memory = deque(maxlen=1000)
self.inventory = pd.DataFrame(columns=['Price', 'POS', 'Order'])
self.is_eval = is_eval
self.learning_rate = env.learning_rate
self.gamma = env.gamma
self.env = env
self.up = dict(batch_size = self.env.batch_size,
frequency = self.env.batch_size)
VPGAgent.__init__(self,
states=dict(type='float', shape=self.state_size.shape),
actions=dict(type='int', num_actions=self.action_size),
network=self.get_network(),
update_mode=self.up,
batching_capacity=self.memory_size,
learning_rate=self.learning_rate,
discount=self.gamma)
self._load_model()
def test_gae(self):
config = Configuration(discount=0.75,
batch_size=8,
learning_rate=0.001,
gae_rewards=True,
gae_lambda=0.5,
states=dict(shape=(1, )),
actions=dict(continuous=True),
network=layered_network_builder(()))
agent = VPGAgent(config=config)
states = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
rewards = [0.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 1.0, 0.0]
terminals = [
False, False, False, False, True, False, False, False, True
]
baseline = np.array([0.25, 0.5, 0.0, 0.25, 0.5, 0.5, 0.25, 0.5, 0.0])
agent.model.baseline = dict(state=Baseline())
agent.model.baseline['state'].predict = lambda states: baseline
td_residuals = np.array([
0.75 * 0.5 - 0.25, 1.0 - 0.5, 0.75 * 0.25, 0.75 * 0.5 - 0.25, 1.0,
1.0 + 0.75 * 0.25 - 0.5, 0.75 * 0.5 - 0.25, 1.0 - 0.5, 0.0
])
result, _ = agent.model.reward_estimation(states=dict(state=states),
rewards=rewards,
terminals=terminals)
expected = np.array([
np.sum(
((0.5 * 0.75)**np.array([0, 1, 2, 3, 4])) * td_residuals[:5]),
np.sum(((0.5 * 0.75)**np.array([0, 1, 2, 3])) * td_residuals[1:5]),
np.sum(((0.5 * 0.75)**np.array([0, 1, 2])) * td_residuals[2:5]),
np.sum(((0.5 * 0.75)**np.array([0, 1])) * td_residuals[3:5]),
np.sum(((0.5 * 0.75)**np.array([0])) * td_residuals[4:5]),
np.sum(((0.5 * 0.75)**np.array([0, 1, 2, 3])) * td_residuals[5:]),
np.sum(((0.5 * 0.75)**np.array([0, 1, 2])) * td_residuals[6:]),
np.sum(((0.5 * 0.75)**np.array([0, 1])) * td_residuals[7:]),
np.sum(((0.5 * 0.75)**np.array([0])) * td_residuals[8:])
])
self.assertTrue((result == expected).all())
for i in range(int(len(infrastructure.peers) * .1))
},
network=[
dict(type='flatten'),
dict(type='dense', size=32, activation='relu'),
],
)
# Create a Vanilla Policy Gradient
agent_vpg = VPGAgent(
states={
"type": 'float',
"shape": infrastructure.get_state().shape
},
actions={
str(i): dict(type="int", num_actions=len(infrastructure.peers))
for i in range(int(len(infrastructure.peers) * .1))
},
network=[
dict(type='flatten'),
dict(type='dense', size=32, activation='relu'),
],
)
#agent_ppo.restore_model("results/client-server")
print("agents made")
monkey = []
rl_ppo = []
rl_dqn = []
rl_vpg = []
# DistributionModel
distributions_spec=None,
entropy_regularization=0.01,
# PGModel
baseline_mode=None,
baseline=None,
baseline_optimizer=None,
gae_lambda=None,
# PGLRModel
likelihood_ratio_clipping=0.2,
summary_spec=None,
distributed_spec=None
)'''
agent = VPGAgent(states_spec=env.states,
actions_spec=env.actions,
network_spec=network_spec,
batch_size=10)
# Create the runner
runner = Runner(agent=agent, environment=env)
# Callback function printing episode statistics
def episode_finished(r):
print(
"Finished episode {ep} after {ts} timesteps (reward: {reward})".format(
ep=r.episode, ts=r.episode_timestep, reward=r.episode_rewards[-1]))
return True
# Start learning
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-r', '--relation', help="Number of episodes")
parser.add_argument('-e',
'--episodes',
type=int,
default=500,
help="Number of episodes")
parser.add_argument('-a',
'--agent',
type=str,
default='vpg',
help="VPG or DQN Agent")
parser.add_argument('-D',
'--debug',
action='store_true',
default=False,
help="Show debug outputs")
args = parser.parse_args()
print("Running DeepPath-TensorForce")
if args.relation: # relation is defined
relation = args.relation
logger.info('Relation set to %s', relation)
else:
logger.error("Error : No Relation name provided!")
return
graphPath = dataPath + 'tasks/' + relation + '/' + 'graph.txt'
relationPath = dataPath + 'tasks/' + relation + '/' + 'train_pos'
if not os.path.exists(relationPath):
logger.info('Incorrect relation specified %s', relation)
print('Incorrect relation specified ', relation)
f = open(relationPath)
data = f.readlines()
f.close()
# Initialize the DeePath Environment class
environment = DPEnv(graphPath, relationPath, task=data)
network_spec = [
dict(type='dense', size=512, activation='relu'),
dict(type='dense', size=1024, activation='relu')
]
step_optimizer = dict(type='adam', learning_rate=1e-3)
agent = None
if args.agent == 'vpg':
logger.info('Initializing VPGAgent')
agent = VPGAgent(states_spec=dict(shape=state_dim, type='float'),
actions_spec=dict(num_actions=action_space,
type='int'),
network_spec=network_spec,
optimizer=step_optimizer,
discount=0.99,
batch_size=1000)
elif args.agent == 'dqn':
logger.info('Initializing DQNAgent')
agent = DQNAgent(states_spec=dict(shape=state_dim, type='float'),
actions_spec=dict(num_actions=action_space,
type='int'),
network_spec=network_spec,
optimizer=step_optimizer,
discount=0.99,
batch_size=1000)
logger.info('Initializing Runner')
runner = Runner(agent=agent, environment=environment)
report_episodes = args.episodes / 50 # default episodes = 500
def episode_finished(r):
if r.episode % report_episodes == 0:
logger.info(
"Finished episode {ep} after {ts} timesteps. Steps Per Second "
.format(ep=r.episode, ts=r.timestep))
logger.info("Episode reward: {}".format(r.episode_rewards[-1]))
logger.info("Average of last 50 rewards: {}".format(
sum(r.episode_rewards[-50:]) / 50))
logger.info("Average of last 100 rewards: {}".format(
sum(r.episode_rewards[-100:]) / 100))
return True
logger.info("Starting {agent} for Environment '{env}'".format(
agent=agent, env=environment))
print("Starting {agent} for Environment".format(agent=agent))
runner.run(episodes=args.episodes,
max_episode_timesteps=1,
episode_finished=episode_finished)
logger.info(
"Learning finished. Total episodes: {ep}".format(ep=runner.episode))
print("Learning finished. Total episodes: {ep}".format(ep=runner.episode))
environment.close()
if timestep % save_step == 0:
plt.savefig('data/%s_%f_%d_%d_%d.jpg' %
(model_name, learning_rate, MAXSTEPS, LAYER_1, LAYER_2))
plt.show()
if __name__ == '__main__':
env = OpenAIGym('Acrobot-v1', visualize=False)
LAYER_1 = 128
LAYER_2 = 64
observation = env.reset()
VPG_agent = VPGAgent(
states=dict(type='float', shape=env.states['shape']),
actions=dict(type='int', num_actions=env.actions['num_actions']),
# discrete action space but continuous state space
network=[
dict(type='dense', size=LAYER_1,
activation='relu'), # changed to tanh for best
dict(type='dense', size=LAYER_2, activation='relu')
],
optimizer=dict(type='adam', learning_rate=learning_rate))
# Create a Proximal Policy Optimization agent
PPO_agent = PPOAgent(
states=dict(type='float', shape=env.states['shape']),
actions=dict(type='int', num_actions=env.actions['num_actions']),
# discrete action space but continuous state space
network=[
dict(type='dense', size=LAYER_1,
activation='relu'), #changed to tanh for best
dict(type='dense', size=LAYER_2, activation='relu')
],
self.gym = gym
self.visualize = False
env = ConcatStates(env)
environment = TFOpenAIGymCust('CryptoPortfolioEIIE-v0', env)
env.seed(0)
state = environment.reset()
state, done, reward = environment.execute(env.action_space.sample())
network_spec = [dict(type='dense', size=16), dict(type='dense', size=10)]
agent = VPGAgent(states_spec=environment.states,
actions_spec=environment.actions,
batch_size=20,
network_spec=network_spec,
discount=0.8,
optimizer=dict(type='adam', learning_rate=1e-4))
runner = Runner(agent=agent, environment=environment, repeat_actions=1)
report_episodes = 100
print("Starting {agent} for Environment '{env}'".format(agent=agent,
env=environment))
pvs = []
def episode_finished(r):
if r.episode % report_episodes == 0:
def get_agent(game, agentType):
count = 1
base_path = '.'
checkpointPath = base_path + "/games/agents/" + game + "/" + agentType + "/"
if agentType == "vpg":
agent = VPGAgent(
states=config[game]["states"],
actions=config[game]["actions"],
memory=1000,
network="auto",
)
elif agentType == "ppo":
agent = PPOAgent(
states=config[game]["states"],
actions=config[game]["actions"],
memory=1000,
network="auto",
)
elif agentType == "dqn":
agent = DQNAgent(
states=config[game]["states"],
actions=config[game]["actions"],
memory=1000,
network="auto",
)
if game == "3pd":
try:
agent.restore(directory=checkpointPath, filename=None)
print("restoration successful")
except Exception as e:
agent.initialize()
for x in tqdm(range(1000001)):
testState = np.full(config[game]["states"]["shape"], None)
for i in range(10):
moveA = agent.act(testState)
moveB = agent.act(testState)
moveC = agent.act(testState)
rewards = payoffs(game, [moveA, moveB, moveC])
if i < 9:
agent.observe(reward=rewards[0], terminal=False)
agent.observe(reward=rewards[1], terminal=False)
agent.observe(reward=rewards[2], terminal=False)
else:
agent.observe(reward=rewards[0], terminal=False)
agent.observe(reward=rewards[1], terminal=False)
agent.observe(reward=rewards[2], terminal=True)
testState[i] = [[moveA], [moveB], [moveC]]
if x % 1000 == 0:
# checkpointPath = "../games/agents/" + game + "/" + agentType + "/"
agent.save(directory=checkpointPath, filename=None)
# print("saving successful")
else:
try:
agent.restore(directory=checkpointPath, filename=None)
print("restoration successful")
except Exception as e:
# try:
# checkpointPath = base_path + "/agents/" + game + "/" + agentType + "/"
# agent.restore(directory=checkpointPath, filename=None)
# print("restoration successful after second attempt")
# except Exception as e:
# a = subprocess.check_output("ls games/", shell=True)
# print(a)
# print(os.getcwd(), "vs", subprocess.check_output("pwd", shell=True))
# checkpointPath = "./games/agents/" + game + "/" + agentType + "/"
# print(checkpointPath)
# agent.restore(directory=checkpointPath, filename=None)
# print("restoration successful after third attempt")
agent.initialize()
for x in tqdm(range(count)):
testState = np.full(config[game]["states"]["shape"], 0)
for i in range(10):
moveA = agent.act(testState)
moveB = agent.act(testState)
rewards = payoffs(game, [moveA, moveB])
if i < 10:
agent.observe(reward=rewards[0], terminal=False)
agent.observe(reward=rewards[1], terminal=False)
else:
agent.observe(reward=rewards[0], terminal=False)
agent.observe(reward=rewards[1], terminal=True)
testState[i] = [[moveA], [moveB]]
checkpointPath = "./games/agents/" + game + "/" + agentType + "/"
agent.save(directory=checkpointPath, filename=None)
print("saving successful")
return agent
network=[dict(type='dense', size=64),
dict(type='dense', size=64)],
batching_capacity=1000,
step_optimizer=dict(type='adam', learning_rate=1e-4))
# Create a Trust Region Policy Optimization agent
agentTRPO = TRPOAgent(
states=dict(type='float', shape=env.observation_space.shape),
actions=dict(type='int', num_actions=env.action_space.n),
network=[dict(type='dense', size=64),
dict(type='dense', size=64)])
# Create a Vanilla Policy Gradient agent
agentVPG = VPGAgent(
states=dict(type='float', shape=env.observation_space.shape),
actions=dict(type='int', num_actions=env.action_space.n),
network=[dict(type='dense', size=64),
dict(type='dense', size=64)])
# Add 3 random agents
agents = []
for agent_id in range(3):
agents.append(SimpleAgent(config["agent"](agent_id, config["game_type"])))
# Add TensorforceAgent
agent_id += 1
agents.append(TensorforceAgent(config["agent"](agent_id, config["game_type"])))
env.set_agents(agents)
env.set_training_agent(agents[-1].agent_id)
env.set_init_game_state(None)
"down": dict(type="float", min_value=0.0, max_value=1.0),
"left": dict(type="float", min_value=0.0, max_value=1.0),
"right": dict(type="float", min_value=0.0, max_value=1.0),
},
network='auto',
memory=10000,
)
elif args.agent == "vpg": # Vanilla Policy Gradient
agent = VPGAgent(
states={
"type": 'float',
"shape": (1, 610)
},
actions={
"up": dict(type="float", min_value=0.0, max_value=1.0),
"down": dict(type="float", min_value=0.0, max_value=1.0),
"left": dict(type="float", min_value=0.0, max_value=1.0),
"right": dict(type="float", min_value=0.0, max_value=1.0),
},
network='auto',
memory=10000,
)
else:
print("Available agents: vpg, ppo, dqn")
exit()
print("agent ready", agent)
agent.initialize() # Set up base of agent
try: # Looks to see if a saved model is available and loads it
network=[
dict(type='flatten'),
dict(type="dense", size=32),
],
memory=10000,
)
elif args.agent == "vpg":
agent = VPGAgent(
states={
"type": 'float',
"shape": G.graph.shape
},
actions={
"user": dict(type="int", num_values=G.graph.shape[0]),
"item": dict(type="int", num_values=G.graph.shape[1])
},
network=[
dict(type='flatten'),
dict(type="dense", size=32),
],
memory=10000,
)
elif args.agent == "trpo":
agent = TRPOAgent(
states={
"type": 'float',
"shape": G.graph.shape
},
actions={
"user": dict(type="int", num_values=G.graph.shape[0]),
str(i): dict(type="int", num_actions=infrastructure.servers) for i in range(int(infrastructure.servers * .1))
},
network=[
dict(type='flatten'),
dict(type="dense", size=32),
dict(type="dense", size=32),
dict(type="dense", size=32)
],
)
elif args.monkey == "vpg":
monkey = VPGAgent(
states={"type":'float', "shape": infrastructure.graph.shape },
actions={
str(i): dict(type="int", num_actions=infrastructure.servers) for i in range(int(infrastructure.servers * .1))
},
network=[
dict(type='flatten'),
dict(type="dense", size=32),
dict(type="dense", size=32),
dict(type="dense", size=32)
],
)
if args.manager == "ppo":
manager = PPOAgent(
states={"type":'float', "shape": infrastructure.graph.shape },
actions={
str(i): dict(type="int", num_actions=infrastructure.servers) for i in range(infrastructure.clients)
},
network=[
dict(type='flatten'),
dict(type="dense", size=32),