def test_ddqn():
args = DDQNArgs()
env = gym.make(args.env_name)
agent = DDQNAgent(env, QNet, SimpleNormalizer, args)
agent.load(args.save_dir)
mean_reward = [agent.test_one_episode(True) for _ in range(100)]
print(np.mean(mean_reward))
def train_ddqn():
args = DDQNArgs()
env = gym.make(args.env_name)
agent = DDQNAgent(env, QNet, SimpleNormalizer, args)
pre_best = -1e9
for ep in range(args.max_ep):
agent.train_one_episode()
if ep % args.test_interval == 0:
r = agent.test_model()
if r > pre_best:
pre_best = r
agent.save(args.save_dir)
def main():
np.random.seed(12345678)
n = 10
env = PermutationSorting(n)
state_transformer = OneHotStateTransformer(n)
agent = DDQNAgent(env, state_transformer)
agent.parallel_pretrain(rows=10000, epochs=30)
# agent.load_pretrain_weights()
agent.train(episodes=10000, max_steps=250)
# agent.load_final_weights()
for _ in range(10):
p = np.random.permutation(n)
rl_ans = agent.solve(p)
exact_ans = PermutationExactSolver(n).solve(p)
print(p, '-', 'RL:', rl_ans, ' Exact:', exact_ans)
import numpy as np
from ddqn import DDQNAgent
from utils import plot_learning_curve, make_env
if __name__ == '__main__':
env = make_env('PongNoFrameskip-v4')
best_score = -np.inf
load_checkpoint = False
n_games = 100
agent = DDQNAgent(gamma=0.99,
epsilon=1.0,
lr=0.0001,
input_dims=(env.observation_space.shape),
n_actions=env.action_space.n,
mem_size=50000,
eps_min=0.1,
batch_size=32,
replace=10000,
eps_dec=1e-5,
chkpt_dir='models/',
algo='DDQNAgent',
env_name='PongNoFrameskip-v4')
if load_checkpoint:
agent.load_models()
fname = agent.algo + '_' + agent.env_name + '_lr' + str(agent.lr) +'_' \
+ str(n_games) + 'games'
figure_file = 'plots/' + fname + '.png'
n_steps = 0