def testEvaluationOption(self):
ray.init()
agent = DQNAgent(env="CartPole-v0", config={"evaluation_interval": 2})
r0 = agent.train()
r1 = agent.train()
r2 = agent.train()
r3 = agent.train()
r4 = agent.train()
self.assertTrue("evaluation" in r0)
self.assertTrue("episode_reward_mean" in r0["evaluation"])
self.assertEqual(r0["evaluation"], r1["evaluation"])
self.assertNotEqual(r1["evaluation"], r2["evaluation"])
self.assertEqual(r2["evaluation"], r3["evaluation"])
self.assertNotEqual(r3["evaluation"], r4["evaluation"])
def testTrainCartpoleOffPolicy(self):
register_env(
"test3", lambda _: PartOffPolicyServing(gym.make("CartPole-v0"),
off_pol_frac=0.2))
dqn = DQNAgent(env="test3", config={"exploration_fraction": 0.001})
for i in range(100):
result = dqn.train()
print("Iteration {}, reward {}, timesteps {}".format(
i, result["episode_reward_mean"], result["timesteps_total"]))
if result["episode_reward_mean"] >= 100:
return
raise Exception("failed to improve reward")
def testTrainCartpoleOffPolicy(self):
register_env(
"test3", lambda _: PartOffPolicyServing(
gym.make("CartPole-v0"), off_pol_frac=0.2))
dqn = DQNAgent(env="test3", config={"exploration_fraction": 0.001})
for i in range(100):
result = dqn.train()
print("Iteration {}, reward {}, timesteps {}".format(
i, result["episode_reward_mean"], result["timesteps_total"]))
if result["episode_reward_mean"] >= 100:
return
raise Exception("failed to improve reward")
"num_gpus": 1,
"hiddens": [],
"schedule_max_timesteps": 7500000,
"timesteps_per_iteration": 4000,
"exploration_fraction": 0.8,
"exploration_final_eps": 0.02,
"lr": 1e-3,
"model": {
"custom_model": "parametric",
"custom_options": {}, # extra options to pass to your model
}
})
# Attempt to restore from checkpoint if possible.
if os.path.exists(CHECKPOINT_FILE):
checkpoint_path = open(CHECKPOINT_FILE).read()
print("Restoring from checkpoint path", checkpoint_path)
dqn.restore(checkpoint_path)
# run the new command using the given tracer
# make a report, placing output in the current directory
# Serving and training loop
while True:
print(pretty_print(dqn.train()))
checkpoint_path = dqn.save()
print("Last checkpoint", checkpoint_path)
with open(CHECKPOINT_FILE, "w") as f:
f.write(checkpoint_path)
"timesteps_per_iteration": 200,
"env_config": {
"observation_size": args.observation_size,
"action_size": args.action_size,
},
})
elif args.run == "PG":
agent = PGAgent(
env="srv",
config={
"num_workers": 0,
"env_config": {
"observation_size": args.observation_size,
"action_size": args.action_size,
},
})
# Attempt to restore from checkpoint if possible.
if os.path.exists(args.checkpoint_file):
checkpoint_file = open(args.checkpoint_file).read()
print("Restoring from checkpoint path", checkpoint_file)
agent.restore(checkpoint_file)
# Serving and training loop
while True:
print(pretty_print(agent.train()))
checkpoint_file = agent.save()
print("Last checkpoint", checkpoint_file)
with open(args.checkpoint_file, "w") as f:
f.write(checkpoint_file)