def train_config(self, gpu):
disable_view_window()
c = Config()
# Note: online policy algorithms such as PPO and A2C does not
# work well in Pendulum (reason unknown)
# and MountainCarContinuous (sparse returns)
c.env_name = "CartPole-v0"
c.env = unwrap_time_limit(gym.make(c.env_name))
c.observe_dim = 4
c.action_num = 2
# maximum and minimum of reward value
# since reward is 1 for every step, maximum q value should be
# below 20(reward_future_steps) * (1 + discount ** n_steps) < 40
c.value_max = 40
c.value_min = 0
c.reward_future_steps = 20
c.max_episodes = 1000
c.max_steps = 200
c.replay_size = 100000
# RAINBOW is not very stable (without dueling and noisy linear)
# compared to other DQNs
c.solved_reward = 180
c.solved_repeat = 5
c.device = gpu
return c
def train_config(self):
disable_view_window()
c = Config()
# the cooperative environment environment provided in
# https://github.com/openai/multiagent-particle-envs
c.env_name = "simple_spread"
c.env = create_env(c.env_name)
c.env.discrete_action_input = True
c.agent_num = 3
c.action_num = c.env.action_space[0].n
c.observe_dim = c.env.observation_space[0].shape[0]
# for contiguous tests
c.test_action_dim = 5
c.test_action_range = 1
c.test_observe_dim = 5
c.test_agent_num = 3
c.max_episodes = 1000
c.max_steps = 200
c.replay_size = 100000
# from https://github.com/wsjeon/maddpg-rllib/tree/master/plots
# PROBLEM: I have no idea how they calculate the rewards
# I cannot replicate their reward curve
c.solved_reward = -15
c.solved_repeat = 5
return c
def train_config(self):
disable_view_window()
c = Config()
# Note: online policy algorithms such as PPO and A2C does not
# work well in Pendulum (reason unknown)
# and MountainCarContinuous (sparse returns)
c.env_name = "CartPole-v0"
c.env = unwrap_time_limit(gym.make(c.env_name))
c.observe_dim = 4
c.action_num = 2
c.max_episodes = 2000 # the actor learns a little bit slower
c.max_steps = 200
c.replay_size = 10000
c.solved_reward = 150
c.solved_repeat = 5
return c