sgoodfriend's picture
PPO playing QbertNoFrameskip-v4 from https://github.com/sgoodfriend/rl-algo-impls/tree/983cb75e43e51cf4ef57f177194ab9a4a1a8808b
a861765
raw
history blame
655 Bytes
import pytest
import gym.spaces
import numpy as np
from rl_algo_impls.shared.policy.actor_critic import clamp_actions
def test_clamp_actions():
action_space = gym.spaces.Box(low=-1, high=1, shape=(1,))
actions = np.array([-1.5, 0, 1.5])
clamped_actions = clamp_actions(actions, action_space, squash_output=False)
np.testing.assert_array_equal(clamped_actions, np.array([-1, 0, 1]))
action_space = gym.spaces.Box(low=-3, high=2, shape=(1,))
actions = np.array([-1, 0, 1])
clamped_actions = clamp_actions(actions, action_space, squash_output=True)
np.testing.assert_array_equal(clamped_actions, np.array([-3, -0.5, 2]))