def evaluate(n_episodes, rl_prio=True):
agent = None
if rl_prio:
config, run = init_run()
agent = get_agent(config, run)
env = get_env(config, rl=True)
else:
env = get_env(rl=False)
env_renderer = RenderTool(env, screen_width=8800)
returns = []
pcs = []
malfs = []
for _ in tqdm(range(n_episodes)):
obs, _ = env.reset(regenerate_schedule=True, regenerate_rail=True)
if RENDER:
env_renderer.reset()
env_renderer.render_env(show=True,
frames=True,
show_observations=False)
if not obs:
break
steps = 0
ep_return = 0
done = defaultdict(lambda: False)
robust_env = CprFlatlandGymEnv(rail_env=env,
max_nr_active_agents=200,
observation_space=None,
priorizer=NrAgentsSameStart(),
allow_noop=True)
# if rl_prio:
# priorities = prio_agent.compute_actions(obs, explore=False)
# sorted_actions = {k: v for k, v in sorted(priorities.items(), key=lambda item: item[1], reverse=True)}
# sorted_handles = list(sorted_actions.keys())
# else:
sorted_handles = robust_env.priorizer.priorize(handles=list(
obs.keys()),
rail_env=env)
while not done['__all__']:
actions = ShortestPathAgent().compute_actions(obs, env)
robust_actions = robust_env.get_robust_actions(
actions, sorted_handles)
obs, all_rewards, done, info = env.step(robust_actions)
if RENDER:
env_renderer.render_env(show=True,
frames=True,
show_observations=False)
print('.', end='', flush=True)
steps += 1
ep_return += np.sum(list(all_rewards.values()))
pc = np.sum(np.array([1 for a in env.agents if is_done(a)
])) / env.get_num_agents()
print("EPISODE PC:", pc)
n_episodes += 1
pcs.append(pc)
returns.append(ep_return /
(env._max_episode_steps * env.get_num_agents()))
malfs.append(
np.sum([a.malfunction_data['nr_malfunctions']
for a in env.agents]))
return pcs, returns, malfs
def evaluate(n_episodes):
run = SUBMISSIONS["rlpr-tcpr"]
config, run = init_run(run)
prio_agent = get_agent(config, run)
run = SUBMISSIONS["rlps-tcpr"]
config, run = init_run(run)
step_agent = ShortestPathRllibAgent(get_agent(config, run), explore=False)
env = get_env(None, rl=True)
env_renderer = RenderTool(env, screen_width=8800)
returns = []
pcs = []
malfs = []
for _ in tqdm(range(n_episodes)):
obs, _ = env.reset(regenerate_schedule=True, regenerate_rail=True)
if RENDER:
env_renderer.reset()
env_renderer.render_env(show=True,
frames=True,
show_observations=False)
if not obs:
break
steps = 0
ep_return = 0
done = defaultdict(lambda: False)
robust_env = CprFlatlandGymEnv(rail_env=env,
max_nr_active_agents=200,
observation_space=None,
priorizer=DistToTargetPriorizer(),
allow_noop=True)
meta_obs = {h: o[1] for h, o in obs.items()}
priorities = prio_agent.compute_actions(meta_obs, explore=False)
sorted_actions = {
k: v
for k, v in sorted(
priorities.items(), key=lambda item: item[1], reverse=True)
}
sorted_handles = list(sorted_actions.keys())
while not done['__all__']:
agent_obs = {h: o[0] for h, o in obs.items()}
actions = step_agent.compute_actions(agent_obs, env)
robust_actions = robust_env.get_robust_actions(
actions, sorted_handles)
obs, all_rewards, done, info = env.step(robust_actions)
if RENDER:
env_renderer.render_env(show=True,
frames=True,
show_observations=False)
print('.', end='', flush=True)
steps += 1
ep_return += np.sum(list(all_rewards.values()))
pc = np.sum(np.array([1 for a in env.agents if is_done(a)
])) / env.get_num_agents()
print("EPISODE PC:", pc)
n_episodes += 1
pcs.append(pc)
returns.append(ep_return /
(env._max_episode_steps * env.get_num_agents()))
malfs.append(
np.sum([a.malfunction_data['nr_malfunctions']
for a in env.agents]))
return pcs, returns, malfs