def test_simple_asymm_ghost_fails(action_sizes):
# Make opponent for asymmetric case
brain_name_opp = BRAIN_NAME + "Opp"
env = SimpleEnvironment(
[BRAIN_NAME + "?team=0", brain_name_opp + "?team=1"],
action_sizes=action_sizes)
# This config should fail because the team that us not learning when both have reached
# max step should be executing the initial, untrained poliy.
self_play_settings = SelfPlaySettings(
play_against_latest_model_ratio=0.0,
save_steps=5000,
swap_steps=5000,
team_change=2000,
)
config = attr.evolve(PPO_TORCH_CONFIG,
self_play=self_play_settings,
max_steps=3000)
check_environment_trains(env, {
BRAIN_NAME: config,
brain_name_opp: config
},
success_threshold=None)
processed_rewards = [
default_reward_processor(rewards)
for rewards in env.final_rewards.values()
]
success_threshold = 0.9
assert any(reward > success_threshold
for reward in processed_rewards) and any(
reward < success_threshold for reward in processed_rewards)
def test_simple_ghost_fails(action_sizes):
env = SimpleEnvironment(
[BRAIN_NAME + "?team=0", BRAIN_NAME + "?team=1"], action_sizes=action_sizes
)
# This config should fail because the ghosted policy is never swapped with a competent policy.
# Swap occurs after max step is reached.
self_play_settings = SelfPlaySettings(
play_against_latest_model_ratio=1.0, save_steps=2000, swap_steps=4000
)
config = attr.evolve(PPO_TORCH_CONFIG, self_play=self_play_settings, max_steps=2500)
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=None)
processed_rewards = [
default_reward_processor(rewards) for rewards in env.final_rewards.values()
]
success_threshold = 0.9
assert any(reward > success_threshold for reward in processed_rewards) and any(
reward < success_threshold for reward in processed_rewards
)