def evaluate(name, environment, n=0, uncertainty=0):
disablePrint()
name = name + '-' + str(n)
agent = pickle.load(
open(f"outputs/{'-'.join(name.split('-')[:-1])}/Agents/{name}.agent",
"rb"))
agent.explore = agent.exploration.greedy
agent.uncertainty = True
agent.state_avoidance = False
agent.uncertainty_weight = uncertainty
env = Environments(render=False,
envs=[environment for _ in range(20)],
agent=agent)
rews, dones = [], []
for i in range(20000):
obs, hn, cn = env.start()
act, obs_old, h0, c0, hn, cn, _, _ = agent.chooseMulti(obs, hn, cn)
obs, rew, done, info = env.step(act, hn, cn)
rews.append(sum(rew))
dones.append(sum(done))
enablePrint()
score = sum(rews) / sum(dones)
print(
name.ljust(20, ' '),
int(100 * (score - rMin[environment]) /
(rMax[environment] - rMin[environment])) / 100, " ", score)
def main():
name, environment, hours, total_agents, done = params['name'], params['environment'], params['hours'], params['total_agents'], None
agent = Agent(**params)
env = Environments(render=False, envs=[environment for _ in range(total_agents)], agent=agent)
collector = Collector(**params)
tid, f = time() + 3600 * hours - 300, 0
while time() < tid:
f += 1
obs, hn, cn = env.start()
act, obs_old, h0, c0, hn, cn, before_trace, after_trace = agent.chooseMulti(obs, hn, cn, done=done)
obs, rew, done, info = env.step(act, hn, cn)
collector.collect(rew, done, act, agent.onpolicy)
if not agent.onpolicy and f > 10:
agent.rememberMulti(obs_old, act, obs, rew, h0, c0, hn, cn, done, before_trace, after_trace)
agent.learn()
saveAgent(agent, name)
saveCollector(collector, name)
env = Environments(render=True,
envs=[
'bigfish', 'bossfight', 'caveflyer', 'chaser',
'climber', 'coinrun', 'dodgeball', 'fruitbot', 'heist',
'jumper', 'leaper', 'maze', 'miner', 'ninja', 'plunder',
'starpilot'
],
agent=agent)
network = NetWork().to(device)
print("Number of parameters in network:", count_parameters(network))
print("Number of parameters in encoder:", count_parameters(network.encoder))
print("Number of parameters in decoder:", count_parameters(network.decoder))
criterion = MSELoss()
optimizer = Adam(network.parameters(), lr=1e-4, weight_decay=1e-5)
for f in range(0, 10000000):
obs, hn, cn = env.start()
act, obs_old, h0, c0, hn, cn = agent.chooseMulti(obs, hn, cn)
obs, rew, done, info = env.step(act, hn, cn)
agent.rememberMulti(obs_old, act, obs, rew, h0, c0, hn, cn, done)
obs, action, obs_next, reward, h0, c0, hn, sn, done = agent.memory.sample(
256)
guess = network(obs)
loss = criterion(guess, obs)
loss.backward()
optimizer.step()
optimizer.zero_grad()
if showPrint:
plt.close('all')
imageBig(obs[0].cpu(), y=200, x=600)
imageBig(guess[0].detach().cpu().clamp(-1, 1), y=200, x=1200)
showPrint = False
