writer = SummaryWriter(log_dir=log_dir)
config = DefaultMunch()
config.seed = seed
config.n_episodes = 40000
config.max_t = 1000
config.buffer_size = 100000
config.batch_size = 200
config.gamma = 0.99
config.tau = 0.001
config.lr_actor = 0.0001
config.lr_critic = 0.0001
config.n_agents = n_agents
config.state_size = state_size * state_multiplier
config.action_size = action_size
config.learn_start = 10000
config.max_action = 1 # maximum value allowed for each action
config.memory = ExperienceReplayMemory(config.buffer_size, seed)
config.update_every = 2
config.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
config_file = open(os.path.join(log_dir, "config.json"), "w+")
config_file.write(
json.dumps(json.loads(
jsonpickle.encode(config, unpicklable=False, max_depth=1)),
indent=4,
sort_keys=True))
config_file.close()
scores = []
scores_std = []
scores_avg = []
rand_seed = 0
config = DefaultMunch()
config.seed = seed
config.n_episodes = 10
config.max_t = 1000
config.buffer_size = 100000
config.batch_size = 200
config.gamma = 0.99
config.tau = 0.001
config.lr_actor = 0.0001
config.lr_critic = 0.001
config.n_agents = n_agents
config.state_size = state_size * state_multiplier
config.action_size = action_size
config.learn_start = 3000
config.max_action = 1
config.memory = ExperienceReplayMemory(config.buffer_size, rand_seed)
config.update_every = 2
config.device = device
rand_seed = 0
scores = []
scores_std = []
scores_avg = []
scores_window = deque(maxlen=100)
agent = MultiAgent(config)
agent.load("./save/checkpoint_success.pth")
global_steps = 0
noise_scheduler = config.noise_scheduler
for i_episode in range(config.n_episodes):
env_info = env.reset(train_mode=False)[env.brain_names[0]]