# Configured in hyperparams/a2c.yml
use_sde = True
n_envs = 4
env = make_vec_env(env_id, n_envs=n_envs, seed=seed)
model = a2c(env,
hyper,
'MlpPolicy',
verbose=verbose,
tensorboard_log=tensorboard_log,
seed=seed,
use_sde=use_sde)
model.learn(total_timesteps=total_timesteps)
custom_eval(model,
env_id,
algo,
seed=seed,
outdir=outdir,
value=hyper["value"])
## Compare to vanilla default execution
model = A2C('MlpPolicy',
env,
verbose=verbose,
tensorboard_log=tensorboard_log,
seed=seed,
use_sde=use_sde)
model.learn(total_timesteps=total_timesteps)
custom_eval(model, env_id, algo, seed, "vanilla")
'params_n_epochs': 20,
'params_n_steps': 32,
'params_net_arch': 'medium',
'params_sde_sample_freq': 64,
'params_vf_coef': 0.261185,
'value': 51.3641
}
env = make_vec_env(env_id, n_envs=4, seed=seed)
model = ppo(env,
hyper,
'MlpPolicy',
verbose=0,
tensorboard_log=tensorboard_log,
seed=seed,
use_sde=True,
device="cpu")
model.learn(total_timesteps=60000)
custom_eval(model,
env_id,
algo,
seed=seed,
outdir="results",
value=hyper["value"])
## Compare to vanilla default execution. Vanilla is no action noise, but tuning always uses action noise(?)
#model = PPO('MlpPolicy', env, verbose = 0, tensorboard_log = tensorboard_log,
# seed = seed, use_sde = use_sde)
#model.learn(total_timesteps = total_timesteps)
#custom_eval(model, env_id, algo, seed, "vanilla")