def test_agent_stats_partial_fit_and_tuple_env():
# Define train and evaluation envs
train_env = (GridWorld, None
) # tuple (constructor, kwargs) must also work in AgentStats
# Parameters
params = {"n_episodes": 500}
horizon = 20
# Run AgentStats
stats = AgentStats(DummyAgent,
train_env,
init_kwargs=params,
n_fit=4,
eval_horizon=10)
stats2 = AgentStats(DummyAgent,
train_env,
init_kwargs=params,
n_fit=4,
eval_horizon=10)
# set some writers
stats.set_writer(0, None)
stats.set_writer(3, None)
# Run partial fit
stats.partial_fit(0.1)
stats.partial_fit(0.5)
for agent in stats.fitted_agents:
assert agent.fraction_fitted == 0.6
for _ in range(2):
stats.partial_fit(0.5)
for agent in stats.fitted_agents:
assert agent.fraction_fitted == 1.0
# Run fit
stats2.fit()
# learning curves
plot_episode_rewards([stats], cumulative=True, show=False)
# compare final policies
compare_policies([stats], eval_horizon=horizon, n_sim=10, show=False)
def test_agent_stats_partial_fit():
# Define train and evaluation envs
train_env = GridWorld()
eval_env = GridWorld()
# Parameters
params = {"n_episodes": 500}
horizon = 20
# Check DummyAgent
agent = DummyAgent(train_env, **params)
agent.fit()
agent.policy(None)
# Run AgentStats
stats = AgentStats(DummyAgent,
train_env,
init_kwargs=params,
n_fit=4,
eval_horizon=10)
# Run partial fit
stats.partial_fit(0.1)
stats.partial_fit(0.5)
for agent in stats.fitted_agents:
assert agent.fraction_fitted == 0.6
for _ in range(2):
stats.partial_fit(0.5)
for agent in stats.fitted_agents:
assert agent.fraction_fitted == 1.0
# learning curves
plot_episode_rewards([stats], cumulative=True, show=False)
# compare final policies
compare_policies([stats],
eval_env,
eval_horizon=horizon,
n_sim=10,
show=False)
from rlberry.agents.ppo import PPOAgent
from rlberry.envs.benchmarks.ball_exploration import PBall2D
from rlberry.seeding import seeding
from rlberry.stats import AgentStats, plot_episode_rewards, compare_policies
seeding.set_global_seed(1223)
env = PBall2D()
n_episodes = 400
horizon = 100
ppo_params = {}
ppo_params['n_episodes'] = 400
ppo_params['horizon'] = 100
ppo_params['gamma'] = 0.99
ppo_params['learning_rate'] = 0.001
ppo_params['eps_clip'] = 0.2
ppo_params['k_epochs'] = 4
ppo_stats = AgentStats(PPOAgent,
env,
eval_horizon=100,
init_kwargs=ppo_params,
n_fit=2)
ppo_stats.partial_fit(0.3)
plot_episode_rewards([ppo_stats], show=False, cumulative=True)
compare_policies([ppo_stats], show=False)
ppo_stats.partial_fit(0.2)
plot_episode_rewards([ppo_stats], show=False, cumulative=True)
compare_policies([ppo_stats], show=True)