epsilon=epsilon,
beta=beta)
optimizer.zero_grad()
L.backward()
optimizer.step()
del L
# the clipping parameter reduces as time goes on
epsilon *= .999
# the regulation term also reduces
# this reduces exploration in later runs
beta *= .995
# get the average reward of the parallel environments
mean_rewards.append(np.mean(total_rewards))
# display some progress every 20 iterations
if (e + 1) % 20 == 0:
print("Episode: {0:d}, score: {1:f}".format(e + 1,
np.mean(total_rewards)))
print(total_rewards)
# update progress widget bar
timer.update(e + 1)
timer.finish()
pong_utils.play(env, policy, time=200)
torch.save(policy, 'PPO.policy')
import pong_utils
device = pong_utils.device
print("using device: ", device)
import gym
env = gym.make('PongDeterministic-v4')
print("List of available actions: ", env.unwrapped.get_action_meanings())
# The actions 'RIGHTFIRE' = 4 and 'LEFTFIRE" = 5 makes the game restarts if done
import matplotlib.pyplot as plt
from agent import Policy
agent = Policy()
agent = agent.to(device)
pong_utils.play(env, agent, time=100)
envs = pong_utils.parallelEnv('PongDeterministic-v4', n=4, seed=12345)
prob, state, action, reward = pong_utils.collect_trajectories(envs,
agent,
tmax=100)