def worker(runs, env_cfg, task, param):
environment = Maze(**env_cfg)
agent_program = AGENT_PROGRAM[task[0]](environment.act_spec,
environment.obs_spec, **task[1])
agent = core.Agent(agent_program, lambda env:
(env.STATE_IDX[env.obs], env.reward),
lambda action, env: env.step(action))
steps = []
for _ in range(0, task[2]['episodes']):
core.Run(agent, environment).start()
steps.append(environment.nstep)
agent_program.reset()
environment.reset()
key = {
**env_cfg,
**{
'runs': runs
},
**{
'alg': task[0]
},
**task[1],
**task[2]
}
SharedMem.dump(Testbed.key_for('episodes', **key), param['run'], steps)
def worker(runs, env_cfg, task, param):
environment = Maze(**env_cfg)
agent_program = AGENT_PROGRAM[task[0]](environment.act_spec,
environment.obs_spec, **task[1])
agent = core.Agent(agent_program, lambda env:
(env.STATE_IDX[env.obs], env.reward),
lambda action, env: env.step(action))
while environment.steps_cnt < task[2]['steps']:
core.Run(agent, environment).start()
agent_program.reset()
environment.reset()
key = {
**env_cfg,
**{
'runs': runs
},
**{
'alg': task[0]
},
**task[1],
**task[2]
}
SharedMem.dump(Testbed.key_for('rewards', **key), param['run'],
environment.rewards[0:task[2]['steps']])
def test_runs_in_mountain_car_environment():
environment = MountainCar()
agent_program = SemiGradientSarsa(
environment.act_spec,
environment.obs_spec,
alpha=0.5,
epsilon=0.0,
gamma=1.0)
agent = core.Agent(
agent_program,
lambda env: (env.obs, env.reward, env.done()),
lambda action, env: env.step(action))
core.Run(agent, environment).start()
def test_true_sarsa_lambda_runs_in_mountain_car_environment():
environment = MountainCar()
agent_program = TrueOnlineSarsaLambda(
environment.act_spec,
environment.obs_spec,
alpha=0.2,
epsilon=0.0,
gamma=1.0,
lmbda=0.9)
agent = core.Agent(
agent_program,
lambda env: (env.obs, env.reward, env.done()),
lambda action, env: env.step(action))
core.Run(agent, environment).start()
def test_collects_reward_from_each_step(self):
arms = 10
eps = 0.1
environment = NArmedBanditEnv(10, arms)
agent = core.Agent(
epsilongreedy.SampleAverage(
act_spec=core.Spec([core.Space(shape=(10, ))]),
obs_spec=core.Spec([core.Space(shape=(10, ))]),
epsilon=eps), lambda env: (env.last_action, env.reward),
lambda action, env: env.step(action))
core.Run(agent, environment).start()
assert len(environment.all_rewards) == 10
assert len(environment.optimal_actions) == 10
def test_nstep_sarsa():
environment = WindyGridWorld()
agent_program = OnPolicyNStepSarsa(
environment.act_spec,
environment.obs_spec,
n=2,
alpha=0.5,
epsilon=0.1,
gamma=1.0)
agent = core.Agent(
agent_program,
lambda env: (env.STATE_IDX[env.obs], env.reward, env.done()),
lambda action, env: env.step(action))
core.Run(agent, environment).start()
def test_actor_critic_runs_in_mountain_car_environment():
environment = MountainCar()
agent_program = ActorCriticLambda(
environment.act_spec,
environment.obs_spec,
alpha_w=0.1,
alpha_theta=0.01,
gamma=1.0,
lambda_w=0.9,
lambda_theta=0.9)
agent = core.Agent(
agent_program,
lambda env: (env.obs, env.reward, env.done()),
lambda action, env: env.step(action))
core.Run(agent, environment).start()
def worker(runs, env_cfg, task, param):
environment = NArmedBanditEnv(**env_cfg)
agent = core.Agent(
AGENT_PROGRAM[task[0]](environment.act_spec, environment.obs_spec,
**task[1]), lambda env:
(env.last_action, env.reward),
lambda action, env: env.step(action))
core.Run(agent, environment).start()
key = {**env_cfg, **{'runs': runs}, **{'alg': task[0]}, **task[1]}
SharedMem.dump(Testbed.key_for('rewards', **key), param['run'],
environment.all_rewards)
SharedMem.dump(Testbed.key_for('actions', **key), param['run'],
environment.optimal_actions)
def worker(runs, env_cfg, task, param):
environment = RaceTrack(**env_cfg)
agent = core.Agent(
AGENT_PROGRAM[task[0]](
environment.act_spec,
environment.obs_spec,
**task[1]),
lambda env: (env.obs, -1),
lambda action, env: env.step(action))
core.Run(agent, environment).start()
key = {**env_cfg, **{'runs': runs}, **{'alg': task[0]}, **task[1]}
states, actions, rewards = environment.episode()
run = param['run']
SharedMem.dump(Testbed.key_for('states', **key), run, states)
SharedMem.dump(Testbed.key_for('rewards', **key), run, rewards)
SharedMem.dump(Testbed.key_for('actions', **key), run, actions)
def run(self):
environment = WindyGridWorld(stochastic=self.stochastic)
agent_program = OnPolicyNStepSarsa(environment.act_spec,
environment.obs_spec,
n=self.n,
alpha=self.alpha,
epsilon=self.epsilon,
gamma=self.gamma)
agent = core.Agent(
agent_program, lambda env:
(env.STATE_IDX[env.obs], env.reward, env.done()),
lambda action, env: env.step(action))
for _ in tqdm.tqdm(range(0, self.runs)):
core.Run(agent, environment).start()
self.steps.append(environment.nstep)
agent_program.reset()
environment.reset()
