if infrastructure.attempts < infrastructure.peers:
agent_dqn.observe(reward=reward, terminal=False)
else:
agent_dqn.observe(reward=reward, terminal=True)
rl_dqn.append(reward)
#trpo agent
action = agent_vpg.act(state)
action = action.values()
reward = infrastructure.monkey(action)
if infrastructure.attempts < infrastructure.peers:
agent_vpg.observe(reward=reward, terminal=False)
else:
agent_vpg.observe(reward=reward, terminal=True)
rl_vpg.append(reward)
#monkey
action = [
randint(1,
len(infrastructure.peers) - 1)
for x in range(int(len(infrastructure.peers) * .2))
]
reward = infrastructure.monkey(action)
#print("monkey", action)
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
model_name = 'VPG'
while timestep < MAXSTEPS and len(durations) <= 100:
action = VPG_agent.act(observation)
observation, done, reward = env.execute(action)
if reward == 0.0:
ep_done = True
else:
ep_done = False
if ep_done:
print(timestep - startstep, ' steps needed to finish this episode')
durations.append(timestep - startstep)
startstep = timestep + 1
# Add experience, agent automatically updates model according to batch size
VPG_agent.observe(reward=reward, terminal=done)
if ep_done:
env.reset() #reinitialize
ep_done = False
agg_sum = agg_sum + reward
agg_rewards.append(agg_sum / timestep)
rewards.append(reward)
if timestep % 1000 == 0:
plotter(timestep,
model_name=model_name,
title='%s Rewards' % model_name,
save_step=1000)
timestep = timestep + 1
if timestep > MAXSTEPS:
print("Timed out. Too many timesteps.")
else: