[sumo-rl] 機械学習で交差点の信号制御を最適化してみた

uv init --app --no-readme --python 3.12 .

uv add "sumo-rl>=1.4.5" eclipse-sumo "stable-baselines3[extra]>=2.3" pandas matplotlib

uv pip list | grep -Ei "sumo|traci"

uv run sumo --version
uv run netconvert --version
SUMO_HOME="$(uv run python -c 'import sumo; print(sumo.SUMO_HOME)')" uv run python -c "import sumo_rl, stable_baselines3"

brew install --cask xquartz

from pathlib import Path

import sumo_rl
from stable_baselines3 import DQN

net_dir = Path(sumo_rl.__file__).parent / "nets" / "single-intersection"

env = sumo_rl.SumoEnvironment(
    net_file=str(net_dir / "single-intersection.net.xml"),
    route_file=str(net_dir / "single-intersection.rou.xml"),
    single_agent=True,
    use_gui=False,
    num_seconds=1000,
)

model = DQN("MlpPolicy", env, verbose=1)
model.learn(total_timesteps=5000)
model.save("dqn_test")

env.close()
print("done: dqn_test.zip を保存しました")

export SUMO_HOME="$(uv run python -c 'import sumo; print(sumo.SUMO_HOME)')"
uv run python minimal_train.py

import argparse
from pathlib import Path

import sumo_rl
from stable_baselines3 import DQN

parser = argparse.ArgumentParser()
parser.add_argument(
    "--random",
    action="store_true",
    help="学習済みモデルを使わず、ランダム行動(学習前相当)で動かす",
)
args = parser.parse_args()

net_dir = Path(sumo_rl.__file__).parent / "nets" / "single-intersection"

env = sumo_rl.SumoEnvironment(
    net_file=str(net_dir / "single-intersection.net.xml"),
    route_file=str(net_dir / "single-intersection.rou.xml"),
    single_agent=True,
    use_gui=True,
    num_seconds=1000,
    additional_sumo_cmd="--delay 100",  # 1ステップ100ms
)
model = None if args.random else DQN.load("dqn_test.zip")

obs, info = env.reset()
done = False
total_reward = 0.0
while not done:
    if model is None:
        action = env.action_space.sample()
    else:
        action, _ = model.predict(obs, deterministic=True)
    obs, reward, terminated, truncated, info = env.step(int(action))
    total_reward += reward
    done = terminated or truncated

mode = "random (学習前相当)" if args.random else "trained (学習後)"
print(f"mode: {mode}")
print("total_reward:", total_reward)
env.close()

// Before
uv run python minimal_gui_check.py --random

// After
uv run python minimal_gui_check.py

export SUMO_HOME="$(uv run python -c 'import sumo; print(sumo.SUMO_HOME)')"
uv run netedit --new

export SUMO_HOME="$(uv run python -c 'import sumo; print(sumo.SUMO_HOME)')"
uv run python "$SUMO_HOME/tools/randomTrips.py" \
  -n path/to/complex.net.xml \
  -r path/to/complex.rou.xml \
  -e 1000 \
  -p 1.5 \
  --fringe-factor 100

uv add supersuit

[project]
name = "sumo-rl-intersection"
version = "0.1.0"
description = "Add your description here"
requires-python = ">=3.12"
dependencies = [
    "eclipse-sumo>=1.27.1",
    "matplotlib>=3.11.2",
    "pandas>=3.0.6",
    "pettingzoo==1.23.1", # ここを追加
    "stable-baselines3[extra]>=2.3",
    "sumo-rl>=1.4.5", 
    "supersuit==3.9.0", # ここを追加
]

uv sync

import supersuit as ss
import sumo_rl
from stable_baselines3 import DQN
from stable_baselines3.common.vec_env import VecMonitor

env = sumo_rl.parallel_env(
    net_file="path/to/complex.net.xml", # 適宜読み替え
    route_file="path/to/complex.rou.xml", # 適宜読み替え
    use_gui=False,
    num_seconds=1000,
)
env = ss.pad_observations_v0(env)  # 信号ごとに観測次元が違うので、大きい方に合わせてパディングする
env = ss.pad_action_space_v0(env)  # 行動空間(フェーズ数)も同様に合わせる
env = ss.pettingzoo_env_to_vec_env_v1(env)
env = ss.concat_vec_envs_v1(env, num_vec_envs=1, num_cpus=1, base_class="stable_baselines3")
env = VecMonitor(env)

model = DQN("MlpPolicy", env, verbose=1)
model.learn(total_timesteps=50000)
model.save("dqn_multiagent_test")

print("done: dqn_multiagent_test.zip を保存しました")