mirror of
https://github.com/vale981/ray
synced 2025-03-12 14:16:39 -04:00
356 lines
16 KiB
Python
356 lines
16 KiB
Python
from abc import ABCMeta
|
|
import logging
|
|
import numpy as np
|
|
import re
|
|
|
|
from ray.rllib.agents.trainer import Trainer
|
|
from ray.rllib.examples.policy.random_policy import RandomPolicy
|
|
from ray.rllib.policy.policy import PolicySpec
|
|
from ray.rllib.utils.annotations import ExperimentalAPI, override
|
|
from ray.rllib.utils.numpy import softmax
|
|
from ray.rllib.utils.typing import TrainerConfigDict, ResultDict
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
@ExperimentalAPI
|
|
class LeagueBuilder(metaclass=ABCMeta):
|
|
def __init__(self, trainer: Trainer, trainer_config: TrainerConfigDict):
|
|
"""Initializes a LeagueBuilder instance.
|
|
|
|
Args:
|
|
trainer: The Trainer object by which this league builder is used.
|
|
Trainer calls `build_league()` after each training step.
|
|
trainer_config: The (not yet validated) config dict to be
|
|
used on the Trainer. Child classes of `LeagueBuilder`
|
|
should preprocess this to add e.g. multiagent settings
|
|
to this config.
|
|
"""
|
|
self.trainer = trainer
|
|
self.config = trainer_config
|
|
|
|
def build_league(self, result: ResultDict) -> None:
|
|
"""Method containing league-building logic. Called after train step.
|
|
|
|
Args:
|
|
result: The most recent result dict with all necessary stats in
|
|
it (e.g. episode rewards) to perform league building
|
|
operations.
|
|
"""
|
|
raise NotImplementedError
|
|
|
|
|
|
@ExperimentalAPI
|
|
class NoLeagueBuilder(LeagueBuilder):
|
|
"""A LeagueBuilder that does nothing.
|
|
|
|
Useful for simple, non-league-building multi-agent setups.
|
|
See e.g.
|
|
`rllib/tuned_examples/alpha_star/multi-agent-cart-pole-alpha-star.yaml`
|
|
"""
|
|
|
|
def build_league(self, result: ResultDict) -> None:
|
|
pass
|
|
|
|
|
|
@ExperimentalAPI
|
|
class AlphaStarLeagueBuilder(LeagueBuilder):
|
|
def __init__(
|
|
self,
|
|
trainer: Trainer,
|
|
trainer_config: TrainerConfigDict,
|
|
num_random_policies: int = 2,
|
|
num_learning_league_exploiters: int = 4,
|
|
num_learning_main_exploiters: int = 4,
|
|
win_rate_threshold_for_new_snapshot: float = 0.8,
|
|
keep_new_snapshot_training_prob: float = 0.0,
|
|
prob_league_exploiter_match: float = 0.33,
|
|
prob_main_exploiter_match: float = 0.33,
|
|
prob_main_exploiter_playing_against_learning_main: float = 0.5,
|
|
):
|
|
"""Initializes a AlphaStarLeagueBuilder instance.
|
|
|
|
Args:
|
|
trainer: The Trainer object by which this league builder is used.
|
|
Trainer calls `build_league()` after each training step.
|
|
trainer_config: The (not yet validated) config dict to be
|
|
used on the Trainer. Child classes of `LeagueBuilder`
|
|
should preprocess this to add e.g. multiagent settings
|
|
to this config.
|
|
num_random_policies: The number of random policies to add to the
|
|
league. This must be an even number (including 0) as these
|
|
will be evenly distributed amongst league- and main- exploiters.
|
|
num_learning_league_exploiters: The number of learning
|
|
league-exploiters to create.
|
|
num_learning_main_exploiters: The number of learning
|
|
main-exploiters to create.
|
|
win_rate_threshold_for_new_snapshot: The win-rate to be achieved
|
|
for a learning policy to get snapshot'd (forked into `self` +
|
|
a new learning or non-learning copy of `self`).
|
|
keep_new_snapshot_training_prob: The probability with which a new
|
|
snapshot should keep training. Note that the policy from which
|
|
this snapshot is taken will continue to train regardless.
|
|
prob_league_exploiter_match: Probability of an episode to become a
|
|
league-exploiter vs snapshot match.
|
|
prob_main_exploiter_match: Probability of an episode to become a
|
|
main-exploiter vs main match.
|
|
prob_main_exploiter_playing_against_learning_main: Probability of
|
|
a main-exploiter vs (training!) main match.
|
|
"""
|
|
super().__init__(trainer, trainer_config)
|
|
|
|
self.win_rate_threshold_for_new_snapshot = win_rate_threshold_for_new_snapshot
|
|
self.keep_new_snapshot_training_prob = keep_new_snapshot_training_prob
|
|
self.prob_league_exploiter_match = prob_league_exploiter_match
|
|
self.prob_main_exploiter_match = prob_main_exploiter_match
|
|
self.prob_main_exploiter_playing_against_learning_main = (
|
|
prob_main_exploiter_playing_against_learning_main
|
|
)
|
|
|
|
assert num_random_policies % 2 == 0, (
|
|
"ERROR: `num_random_policies` must be even number (we'll distribute "
|
|
"these evenly amongst league- and main-exploiters)!"
|
|
)
|
|
|
|
# Build trainer's multiagent config.
|
|
ma_config = self.config["multiagent"]
|
|
# Make sure the multiagent config dict has no policies defined:
|
|
assert not ma_config.get("policies"), (
|
|
"ERROR: `config.multiagent.policies` should not be pre-defined! "
|
|
"AlphaStarLeagueBuilder will construct this itself."
|
|
)
|
|
ma_config["policies"] = policies = {}
|
|
|
|
self.main_policies = 1
|
|
self.league_exploiters = (
|
|
num_learning_league_exploiters + num_random_policies / 2
|
|
)
|
|
self.main_exploiters = num_learning_main_exploiters + num_random_policies / 2
|
|
|
|
# Add 1 initial (learning) main policy.
|
|
policies["main_0"] = PolicySpec()
|
|
|
|
# Train all non-random policies that exist at beginning.
|
|
ma_config["policies_to_train"] = ["main_0"]
|
|
|
|
# Add random policies.
|
|
i = -1
|
|
for i in range(num_random_policies // 2):
|
|
policies[f"league_exploiter_{i}"] = PolicySpec(policy_class=RandomPolicy)
|
|
policies[f"main_exploiter_{i}"] = PolicySpec(policy_class=RandomPolicy)
|
|
# Add initial (learning) league-exploiters.
|
|
for j in range(num_learning_league_exploiters):
|
|
pid = f"league_exploiter_{j + i + 1}"
|
|
policies[pid] = PolicySpec()
|
|
ma_config["policies_to_train"].append(pid)
|
|
# Add initial (learning) main-exploiters.
|
|
for j in range(num_learning_league_exploiters):
|
|
pid = f"main_exploiter_{j + i + 1}"
|
|
policies[pid] = PolicySpec()
|
|
ma_config["policies_to_train"].append(pid)
|
|
|
|
# Initial policy mapping function: main_0 vs main_exploiter_0.
|
|
ma_config["policy_mapping_fn"] = (
|
|
lambda aid, ep, worker, **kw: "main_0"
|
|
if ep.episode_id % 2 == aid
|
|
else "main_exploiter_0"
|
|
)
|
|
|
|
@override(LeagueBuilder)
|
|
def build_league(self, result: ResultDict) -> None:
|
|
local_worker = self.trainer.workers.local_worker()
|
|
|
|
# If no evaluation results -> Use hist data gathered for training.
|
|
if "evaluation" in result:
|
|
hist_stats = result["evaluation"]["hist_stats"]
|
|
else:
|
|
hist_stats = result["hist_stats"]
|
|
|
|
trainable_policies = local_worker.get_policies_to_train()
|
|
non_trainable_policies = (
|
|
set(local_worker.policy_map.keys()) - trainable_policies
|
|
)
|
|
|
|
logger.info(f"League building after iter {self.trainer.iteration}:")
|
|
|
|
# Calculate current win-rates.
|
|
for policy_id, rew in hist_stats.items():
|
|
mo = re.match("^policy_(.+)_reward$", policy_id)
|
|
if mo is None:
|
|
continue
|
|
policy_id = mo.group(1)
|
|
|
|
# Calculate this policy's win rate.
|
|
won = 0
|
|
for r in rew:
|
|
if r > 0.0: # win = 1.0; loss = -1.0
|
|
won += 1
|
|
win_rate = won / len(rew)
|
|
# TODO: This should probably be a running average
|
|
# (instead of hard-overriding it with the most recent data).
|
|
self.trainer.win_rates[policy_id] = win_rate
|
|
|
|
# Policy is a snapshot (frozen) -> Ignore.
|
|
if policy_id not in trainable_policies:
|
|
continue
|
|
|
|
logger.info(f"\t{policy_id} win-rate={win_rate} -> ", end="")
|
|
|
|
# If win rate is good enough -> Snapshot current policy and decide,
|
|
# whether to freeze the new snapshot or not.
|
|
if win_rate >= self.win_rate_threshold_for_new_snapshot:
|
|
is_main = re.match("^main(_\\d+)?$", policy_id)
|
|
|
|
# Probability that the new snapshot is trainable.
|
|
keep_training_p = self.keep_new_snapshot_training_prob
|
|
# For main, new snapshots are never trainable, for all others
|
|
# use `config.keep_new_snapshot_training_prob` (default: 0.0!).
|
|
keep_training = (
|
|
False
|
|
if is_main
|
|
else np.random.choice(
|
|
[True, False], p=[keep_training_p, 1.0 - keep_training_p]
|
|
)
|
|
)
|
|
# New league-exploiter policy.
|
|
if policy_id.startswith("league_ex"):
|
|
new_pol_id = re.sub(
|
|
"_\\d+$", f"_{self.league_exploiters}", policy_id
|
|
)
|
|
self.league_exploiters += 1
|
|
# New main-exploiter policy.
|
|
elif policy_id.startswith("main_ex"):
|
|
new_pol_id = re.sub("_\\d+$", f"_{self.main_exploiters}", policy_id)
|
|
self.main_exploiters += 1
|
|
# New main policy snapshot.
|
|
else:
|
|
new_pol_id = re.sub("_\\d+$", f"_{self.main_policies}", policy_id)
|
|
self.main_policies += 1
|
|
|
|
if keep_training:
|
|
trainable_policies.add(new_pol_id)
|
|
else:
|
|
non_trainable_policies.add(new_pol_id)
|
|
|
|
logger.info(
|
|
f"adding new opponents to the mix ({new_pol_id}; "
|
|
f"trainable={keep_training})."
|
|
)
|
|
|
|
num_main_policies = self.main_policies
|
|
probs_match_types = [
|
|
self.prob_league_exploiter_match,
|
|
self.prob_main_exploiter_match,
|
|
1.0
|
|
- self.prob_league_exploiter_match
|
|
- self.prob_main_exploiter_match,
|
|
]
|
|
prob_playing_learning_main = (
|
|
self.prob_main_exploiter_playing_against_learning_main
|
|
)
|
|
|
|
# Update our mapping function accordingly.
|
|
def policy_mapping_fn(agent_id, episode, worker, **kwargs):
|
|
|
|
# Pick, whether this is:
|
|
# LE: league-exploiter vs snapshot.
|
|
# ME: main-exploiter vs (any) main.
|
|
# M: Learning main vs itself.
|
|
type_ = np.random.choice(["LE", "ME", "M"], p=probs_match_types)
|
|
|
|
# Learning league exploiter vs a snapshot.
|
|
# Opponent snapshots should be selected based on a win-rate-
|
|
# derived probability.
|
|
if type_ == "LE":
|
|
if episode.episode_id % 2 == agent_id:
|
|
league_exploiter = np.random.choice(
|
|
[
|
|
p
|
|
for p in trainable_policies
|
|
if p.startswith("league_ex")
|
|
]
|
|
)
|
|
logger.debug(
|
|
f"Episode {episode.episode_id}: AgentID "
|
|
f"{agent_id} played by {league_exploiter} (training)"
|
|
)
|
|
return league_exploiter
|
|
# Play against any non-trainable policy (excluding itself).
|
|
else:
|
|
all_opponents = list(non_trainable_policies)
|
|
probs = softmax(
|
|
[
|
|
worker.global_vars["win_rates"][pid]
|
|
for pid in all_opponents
|
|
]
|
|
)
|
|
opponent = np.random.choice(all_opponents, p=probs)
|
|
logger.debug(
|
|
f"Episode {episode.episode_id}: AgentID "
|
|
f"{agent_id} played by {opponent} (frozen)"
|
|
)
|
|
return opponent
|
|
|
|
# Learning main exploiter vs (learning main OR snapshot main).
|
|
elif type_ == "ME":
|
|
if episode.episode_id % 2 == agent_id:
|
|
main_exploiter = np.random.choice(
|
|
[
|
|
p
|
|
for p in trainable_policies
|
|
if p.startswith("main_ex")
|
|
]
|
|
)
|
|
logger.debug(
|
|
f"Episode {episode.episode_id}: AgentID "
|
|
f"{agent_id} played by {main_exploiter} (training)"
|
|
)
|
|
return main_exploiter
|
|
else:
|
|
# n% of the time, play against the learning main.
|
|
# Also always play againt learning main if no
|
|
# non-learning mains have been created yet.
|
|
if num_main_policies == 1 or (
|
|
np.random.random() < prob_playing_learning_main
|
|
):
|
|
main = "main_0"
|
|
training = "training"
|
|
# 100-n% of the time, play against a non-learning
|
|
# main. Opponent main snapshots should be selected
|
|
# based on a win-rate-derived probability.
|
|
else:
|
|
all_opponents = [
|
|
f"main_{p}"
|
|
for p in list(range(1, num_main_policies))
|
|
]
|
|
probs = softmax(
|
|
[
|
|
worker.global_vars["win_rates"][pid]
|
|
for pid in all_opponents
|
|
]
|
|
)
|
|
main = np.random.choice(all_opponents, p=probs)
|
|
training = "frozen"
|
|
logger.debug(
|
|
f"Episode {episode.episode_id}: AgentID "
|
|
f"{agent_id} played by {main} ({training})"
|
|
)
|
|
return main
|
|
|
|
# Main policy: Self-play.
|
|
else:
|
|
logger.debug(f"Episode {episode.episode_id}: main_0 vs main_0")
|
|
return "main_0"
|
|
|
|
# Add and set the weights of the new polic(y/ies).
|
|
state = self.trainer.get_policy(policy_id).get_state()
|
|
self.trainer.add_policy(
|
|
policy_id=new_pol_id,
|
|
policy_cls=type(self.trainer.get_policy(policy_id)),
|
|
policy_state=state,
|
|
policy_mapping_fn=policy_mapping_fn,
|
|
policies_to_train=trainable_policies,
|
|
)
|
|
|
|
else:
|
|
logger.info("not good enough; will keep learning ...")
|