config["timesteps_per_batch"] = HORIZON * N_ROLLOUTS
config["gamma"] = 0.999 # discount rate
config["model"].update({"fcnet_hiddens": [32, 32, 32]})
config["use_gae"] = True
config["lambda"] = 0.97
config["sgd_batchsize"] = min(16 * 1024, config["timesteps_per_batch"])
config["kl_target"] = 0.02
config["num_sgd_iter"] = 10
config["horizon"] = HORIZON
# save the flow params for replay
flow_json = json.dumps(flow_params, cls=FlowParamsEncoder, sort_keys=True,
indent=4)
config['env_config']['flow_params'] = flow_json
create_env, env_name = make_create_env(params=flow_params, version=0)
# Register as rllib env
register_env(env_name, create_env)
trials = run_experiments({
"highway_stabilize": {
"run": "PPO",
"env": env_name,
"config": {
**config
},
"checkpoint_freq": 5,
"max_failures": 999,
"stop": {
"training_iteration": 200,
args = parser.parse_args()
result_dir = args.result_dir if args.result_dir[-1] != '/' \
else args.result_dir[:-1]
config = get_rllib_config(result_dir)
# Run on only one cpu for rendering purposes
ray.init(num_cpus=1)
config["num_workers"] = 1
flow_params = get_flow_params(config)
# Create and register a gym+rllib env
create_env, env_name = make_create_env(params=flow_params,
version=0,
sumo_binary="sumo")
register_env(env_name, create_env)
agent_cls = get_agent_class(args.run)
agent = agent_cls(env=env_name, registry=get_registry(), config=config)
checkpoint = result_dir + '/checkpoint-' + args.checkpoint_num
agent._restore(checkpoint)
# Recreate the scenario from the pickled parameters
exp_tag = flow_params["exp_tag"]
net_params = flow_params['net']
vehicles = flow_params['veh']
initial_config = flow_params['initial']
module = __import__("flow.scenarios", fromlist=[flow_params["scenario"]])
scenario_class = getattr(module, flow_params["scenario"])