def evaluate_ppo_hm_and_bc(layout,
ppo_hm_path,
bc_test_path,
num_rounds,
seeds,
best=False,
display=False):
ppo_hm_performance = defaultdict(lambda: defaultdict(list))
agent_bc_test, bc_params = get_bc_agent_from_saved(bc_test_path)
del bc_params["data_params"]
del bc_params["mdp_fn_params"]
evaluator = AgentEvaluator(**bc_params)
for seed in seeds:
agent_ppo, _ = get_ppo_agent(ppo_hm_path, seed, best=best)
ppo_and_bc = evaluator.evaluate_agent_pair(AgentPair(
agent_ppo, agent_bc_test),
num_games=num_rounds,
display=display)
avg_ppo_and_bc = np.mean(ppo_and_bc['ep_returns'])
ppo_hm_performance[layout]["PPO_HM+BC_test_0"].append(avg_ppo_and_bc)
bc_and_ppo = evaluator.evaluate_agent_pair(AgentPair(
agent_bc_test, agent_ppo),
num_games=num_rounds,
display=display)
avg_bc_and_ppo = np.mean(bc_and_ppo['ep_returns'])
ppo_hm_performance[layout]["PPO_HM+BC_test_1"].append(avg_bc_and_ppo)
return ppo_hm_performance
def evaluate_bc_models(bc_model_paths, num_rounds):
"""
Evaluate BC models passed in over `num_rounds` rounds
"""
best_bc_models_performance = {}
# Evaluate best
for layout_name in bc_model_paths['train'].keys():
print(layout_name)
best_bc_models_performance[layout_name] = {}
eval_trajs = eval_with_benchmarking_from_saved(num_rounds, bc_model_paths['train'][layout_name])
best_bc_models_performance[layout_name]["BC_train+BC_train"] = mean_and_std_err(eval_trajs['ep_returns'])
eval_trajs = eval_with_benchmarking_from_saved(num_rounds, bc_model_paths['test'][layout_name])
best_bc_models_performance[layout_name]["BC_test+BC_test"] = mean_and_std_err(eval_trajs['ep_returns'])
bc_train, bc_params_train = get_bc_agent_from_saved(bc_model_paths['train'][layout_name])
bc_test, bc_params_test = get_bc_agent_from_saved(bc_model_paths['test'][layout_name])
del bc_params_train["data_params"]
del bc_params_test["data_params"]
assert common_keys_equal(bc_params_train, bc_params_test)
ae = AgentEvaluator(mdp_params=bc_params_train["mdp_params"], env_params=bc_params_train["env_params"])
train_and_test = ae.evaluate_agent_pair(AgentPair(bc_train, bc_test), num_games=num_rounds)
best_bc_models_performance[layout_name]["BC_train+BC_test_0"] = mean_and_std_err(train_and_test['ep_returns'])
test_and_train = ae.evaluate_agent_pair(AgentPair(bc_test, bc_train), num_games=num_rounds)
best_bc_models_performance[layout_name]["BC_train+BC_test_1"] = mean_and_std_err(test_and_train['ep_returns'])
return best_bc_models_performance
def test_embedded_planning_agent(self):
agent_evaluator = AgentEvaluator({"layout_name": "cramped_room"},
{"horizon": 100})
other_agent = GreedyHumanModel(agent_evaluator.mlp)
epa = EmbeddedPlanningAgent(other_agent,
agent_evaluator.mlp,
agent_evaluator.env,
delivery_horizon=1)
ap = AgentPair(epa, other_agent)
agent_evaluator.evaluate_agent_pair(ap, num_games=1, display=DISPLAY)
class TestAgentEvaluator(unittest.TestCase):
def setUp(self):
self.agent_eval = AgentEvaluator({"layout_name": "cramped_room"},
{"horizon": 100})
def test_human_model_pair(self):
trajs = self.agent_eval.evaluate_human_model_pair()
try:
AgentEvaluator.check_trajectories(trajs)
except AssertionError as e:
self.fail("Trajectories were not returned in standard format:\n{}".
format(e))
def test_rollouts(self):
ap = AgentPair(RandomAgent(), RandomAgent())
trajs = self.agent_eval.evaluate_agent_pair(ap, num_games=5)
try:
AgentEvaluator.check_trajectories(trajs)
except AssertionError as e:
self.fail("Trajectories were not returned in standard format:\n{}".
format(e))
def test_mlp_computation(self):
try:
self.agent_eval.mlp
except Exception as e:
self.fail("Failed to compute MediumLevelPlanner:\n{}".format(e))
def eval_with_benchmarking_from_model(n_games, model, bc_params, no_waits, display=False):
bc_params = copy.deepcopy(bc_params)
a0 = get_bc_agent_from_model(model, bc_params, no_waits)
a1 = get_bc_agent_from_model(model, bc_params, no_waits)
del bc_params["data_params"], bc_params["mdp_fn_params"]
a_eval = AgentEvaluator(**bc_params)
ap = AgentPair(a0, a1)
trajectories = a_eval.evaluate_agent_pair(ap, num_games=n_games, display=display)
return trajectories
def evaluate_ppo_and_bc_models_for_layout(layout,
num_rounds,
bc_model_paths,
ppo_bc_model_paths,
seeds,
best=False,
display=False):
assert len(seeds["bc_train"]) == len(seeds["bc_test"])
ppo_bc_performance = defaultdict(lambda: defaultdict(list))
agent_bc_test, bc_params = get_bc_agent_from_saved(
bc_model_paths['test'][layout])
ppo_bc_train_path = ppo_bc_model_paths['bc_train'][layout]
ppo_bc_test_path = ppo_bc_model_paths['bc_test'][layout]
evaluator = AgentEvaluator(mdp_params=bc_params["mdp_params"],
env_params=bc_params["env_params"])
for seed_idx in range(len(seeds["bc_train"])):
agent_ppo_bc_train, ppo_config = get_ppo_agent(
ppo_bc_train_path, seeds["bc_train"][seed_idx], best=best)
assert common_keys_equal(bc_params["mdp_params"],
ppo_config["mdp_params"])
# For curiosity, how well does agent do with itself?
# ppo_and_ppo = evaluator.evaluate_agent_pair(AgentPair(agent_ppo_bc_train, agent_ppo_bc_train), num_games=max(int(num_rounds/2), 1), display=display)
# avg_ppo_and_ppo = np.mean(ppo_and_ppo['ep_returns'])
# ppo_bc_performance[layout]["PPO_BC_train+PPO_BC_train"].append(avg_ppo_and_ppo)
# How well it generalizes to new agent in simulation?
ppo_and_bc = evaluator.evaluate_agent_pair(AgentPair(
agent_ppo_bc_train, agent_bc_test),
num_games=num_rounds,
display=display)
avg_ppo_and_bc = np.mean(ppo_and_bc['ep_returns'])
ppo_bc_performance[layout]["PPO_BC_train+BC_test_0"].append(
avg_ppo_and_bc)
bc_and_ppo = evaluator.evaluate_agent_pair(AgentPair(
agent_bc_test, agent_ppo_bc_train),
num_games=num_rounds,
display=display)
avg_bc_and_ppo = np.mean(bc_and_ppo['ep_returns'])
ppo_bc_performance[layout]["PPO_BC_train+BC_test_1"].append(
avg_bc_and_ppo)
# How well could we do if we knew true model BC_test?
agent_ppo_bc_test, ppo_config = get_ppo_agent(
ppo_bc_test_path, seeds["bc_test"][seed_idx], best=best)
assert common_keys_equal(bc_params["mdp_params"],
ppo_config["mdp_params"])
ppo_and_bc = evaluator.evaluate_agent_pair(AgentPair(
agent_ppo_bc_test, agent_bc_test),
num_games=num_rounds,
display=display)
avg_ppo_and_bc = np.mean(ppo_and_bc['ep_returns'])
ppo_bc_performance[layout]["PPO_BC_test+BC_test_0"].append(
avg_ppo_and_bc)
bc_and_ppo = evaluator.evaluate_agent_pair(AgentPair(
agent_bc_test, agent_ppo_bc_test),
num_games=num_rounds,
display=display)
avg_bc_and_ppo = np.mean(bc_and_ppo['ep_returns'])
ppo_bc_performance[layout]["PPO_BC_test+BC_test_1"].append(
avg_bc_and_ppo)
return ppo_bc_performance