def run(args):
if args.train_pg:
env_name = args.env_name or 'LunarLander-v2'
env = Environment(env_name, args, atari_wrapper=False)
from agent_dir.agent_pg import AgentPG
agent = AgentPG(env, args)
agent.train()
if args.train_dqn:
env_name = args.env_name or 'MsPacmanNoFrameskip-v0'
env = Environment(env_name, args, atari_wrapper=True)
from agent_dir.agent_dqn import AgentDQN
agent = AgentDQN(env, args)
agent.train()
if args.test_pg:
env = Environment('LunarLander-v2', args, test=True)
from agent_dir.agent_pg import AgentPG
agent = AgentPG(env, args)
test(agent, env, total_episodes=30)
if args.test_dqn:
env = Environment('MsPacmanNoFrameskip-v0',
args,
atari_wrapper=True,
test=True)
from agent_dir.agent_dqn import AgentDQN
agent = AgentDQN(env, args)
test(agent, env, total_episodes=50)
def run(args):
if args.train_pg:
env_name = args.env_name or 'LunarLander-v2'
env = Environment(env_name, args, atari_wrapper=False)
from agent_dir.agent_pg import AgentPG
agent = AgentPG(env, args)
agent.train()
if args.train_pg_improve:
env_name = args.env_name or 'LunarLander-v2'
env = Environment(env_name, args, atari_wrapper=False)
from agent_dir.agent_pg_improve import AgentPG
agent = AgentPG(env, args)
agent.train()
if args.train_dqn:
env_name = args.env_name or 'AssaultNoFrameskip-v0'
env = Environment(env_name, args, atari_wrapper=True)
from agent_dir.agent_dqn import AgentDQN
agent = AgentDQN(env, args)
agent.train()
if args.train_mario:
from agent_dir.agent_mario import AgentMario
agent = AgentMario(None, args)
agent.train()
if args.test_pg:
env = Environment('LunarLander-v2', args, test=True)
from agent_dir.agent_pg import AgentPG
agent = AgentPG(env, args)
test(agent, env, total_episodes=30)
if args.test_pg_improve:
env = Environment('LunarLander-v2', args, test=True)
from agent_dir.agent_pg_improve import AgentPG
agent = AgentPG(env, args)
test(agent, env, total_episodes=30)
if args.test_dqn:
env = Environment('AssaultNoFrameskip-v0',
args,
atari_wrapper=True,
test=True)
from agent_dir.agent_dqn import AgentDQN
agent = AgentDQN(env, args)
test(agent, env, total_episodes=100)
if args.test_mario:
env = Environment('SuperMarioBros-v0', args)
from agent_dir.agent_mario import AgentMario
agent = AgentMario(env, args)
test(agent, env, total_episodes=10)
def run(args):
"******Policy Gradient******"
if args.train_pg:
env_name = args.env_name or 'LunarLander-v2'
env = Environment(env_name, args, atari_wrapper=False)
if args.pg_type == 'pg':
from agent_dir.agent_pg import AgentPG
agent = AgentPG(env, args)
elif args.pg_type == 'pg_nor':
from agent_dir.agent_pg import AgentPG
agent = AgentPG(env, args, normalization=True)
elif args.pg_type == 'pg_ppo':
from agent_dir.agent_pg_ppo import AgentPG
agent = AgentPG(env, args)
agent.train()
if args.test_pg:
env = Environment('LunarLander-v2', args, test=True)
from agent_dir.agent_pg import AgentPG
agent = AgentPG(env, args)
test(agent, env, total_episodes=30)
"******Deep Q Learning******"
if args.train_dqn:
env_name = args.env_name or 'AssaultNoFrameskip-v0'
env = Environment(env_name, args, atari_wrapper=True)
from agent_dir.agent_dqn import AgentDQN
agent = AgentDQN(env, args)
agent.train()
if args.test_dqn:
env = Environment('AssaultNoFrameskip-v0',
args,
atari_wrapper=True,
test=True)
from agent_dir.agent_dqn import AgentDQN
agent = AgentDQN(env, args)
test(agent, env, total_episodes=100)
if args.train_a2c:
from agent_dir.agent_a2c import AgentA2C
agent = AgentA2C(None, args)
agent.train()
if args.test_a2c:
env = Environment('SuperMarioBros-v0', args)
from agent_dir.agent_a2c import AgentA2C
agent = AgentA2C(env, args)
test(agent, env, total_episodes=10)
def run(args):
if args.train_pg:
env_name = args.env_name or 'LunarLander-v2'
env = Environment(env_name, args, atari_wrapper=False)
from agent_dir.agent_pg import AgentPG
agent = AgentPG(env, args)
agent.train()
if args.train_dqn:
env_name = args.env_name or 'MsPacmanNoFrameskip-v0'
env = Environment(env_name, args, atari_wrapper=True)
from agent_dir.agent_dqn import AgentDQN
agent = AgentDQN(env, args)
agent.train()
if args.test_pg:
env = Environment('LunarLander-v2', args, test=True)
from agent_dir.agent_pg import AgentPG
agent = AgentPG(env, args)
test(agent, env, total_episodes=30)
if args.test_dqn:
env = Environment('MsPacmanNoFrameskip-v0', args, atari_wrapper=True, test=True)
from agent_dir.agent_dqn import AgentDQN
agent = AgentDQN(env, args)
test(agent, env, total_episodes=50)
if args.plot_dqn == 1:
import matplotlib.pyplot as plt
env_name = args.env_name or 'MsPacmanNoFrameskip-v0'
env = Environment(env_name, args, atari_wrapper=True)
from agent_dqn_plot import AgentDQN
agent = AgentDQN(env, args)
dqn_training = agent.train()
ret = None
while(ret==None or ret[-1] < 25):
window_size = 20
ret = next(dqn_training)
plt.plot(ret)
plt.xlabel('number of episodes playing')
plt.ylabel('average reward of last {} episodes'.format(window_size))
plt.title('learning curve of dqn with pacman')
plt.savefig('dqn-learning_curve1.png')
plt.close()
elif args.plot_dqn == 2:
import matplotlib.pyplot as plt
game_name = 'MsPacmanNoFrameskip-v0'
#game_name = 'AtlantisNoFrameskip-v0'
#game_name = 'UpNDownNoFrameskip-v0'
#game_name = 'AmidarNoFrameskip-v0'
#game_name = 'GopherNoFrameskip-v0'
#game_name = 'SkiingNoFrameskip-v0'
env_name = args.env_name or game_name
env1 = Environment(env_name, args, atari_wrapper=True)
env2 = Environment(env_name, args, atari_wrapper=True)
env3 = Environment(env_name, args, atari_wrapper=True)
env4 = Environment(env_name, args, atari_wrapper=True)
from agent_dqn_plot import AgentDQN
agent1 = AgentDQN(env1, args, target_update_freq=100000)
agent2 = AgentDQN(env2, args, target_update_freq=10000)
agent3 = AgentDQN(env3, args, target_update_freq=1000)
agent4 = AgentDQN(env4, args, target_update_freq=1)
dqn_training1 = agent1.train()
dqn_training2 = agent2.train()
dqn_training3 = agent3.train()
dqn_training4 = agent4.train()
for episode in range(6000):
window_size = 20
ret1 = next(dqn_training1)
ret2 = next(dqn_training2)
ret3 = next(dqn_training3)
ret4 = next(dqn_training4)
if episode % 10 == 0:
plt.plot(ret1, label='target_update_freq=100000')
plt.plot(ret2, label='target_update_freq=10000')
plt.plot(ret3, label='target_update_freq=1000')
plt.plot(ret4, label='target_update_freq=1')
plt.xlabel('number of episodes playing')
plt.ylabel('average reward of last {} episodes'.format(window_size))
plt.legend()
plt.title('learning curve of dqn with {}'.format(game_name))
plt.savefig('dqn-learning_curve2.png')
plt.close()
#print(ret)
elif args.plot_dqn == 3:
print('Duel vs DQN')
#print('DDQN vs DQN')
import matplotlib.pyplot as plt
#game_name = 'MsPacmanNoFrameskip-v0'
#game_name = 'AlienNoFrameskip-v4'
game_name = 'AtlantisNoFrameskip-v4'
#game_name = 'UpNDownNoFrameskip-v0'
#game_name = 'AsterixNoFrameskip-v0'
#env_name = args.env_name or game_name
env_name = game_name
env1 = Environment(env_name, args, atari_wrapper=True)
env2 = Environment(env_name, args, atari_wrapper=True)
from agent_dqn_plot import AgentDQN
from Duel import AgentDuel
from DDQN import AgentDDQN
agent1 = AgentDQN(env1, args)
agent2 = AgentDuel(env2, args)
#agent2 = AgentDDQN(env2, args)
dqn_training1 = agent1.train()
dqn_training2 = agent2.train()
for episode in range(3500):
window_size = 20
ret1 = next(dqn_training1)
ret2 = next(dqn_training2)
if episode % 10 == 0:
plt.plot(ret1, label='DQN')
#plt.plot(ret2, label='DDQN')
plt.plot(ret2, label='Duel')
plt.xlabel('number of episodes playing')
plt.ylabel('average reward of last {} episodes'.format(window_size))
plt.legend()
plt.title('learning curve of dqn with {}'.format(env_name))
#plt.savefig('dqn-learning_curve_ddqn.png')
plt.savefig(game_name + 'dqn-learning_curve_duel.png')
plt.close()
#print(ret)
if args.plot_pg == 1:
import matplotlib.pyplot as plt
env_name = args.env_name or 'LunarLander-v2'
env = Environment(env_name, args, atari_wrapper=False)
from agent_pg_plot import AgentPG
agent = AgentPG(env, args)
print(agent.model)
pg_training = agent.train()
for _ in range(600):
result = next(pg_training)
"""plot"""
plt.plot(range(len(result)), result, label='policy gradient')
plt.title('learning curve of pg with {}'.format(env_name))
plt.ylabel('average reward of last 20 episodes')
plt.xlabel('number of episodes playing')
plt.legend()
plt.savefig('pg-learning_curve.png')
plt.close()
if args.plot_pg == 3:
import matplotlib.pyplot as plt
env_name = args.env_name or 'LunarLander-v2'
env1 = Environment(env_name, args, atari_wrapper=False)
env2 = Environment(env_name, args, atari_wrapper=False)
from agent_pg_plot import AgentPG
from importance_sampling import Agent
agent1 = AgentPG(env1, args)
agent2 = Agent(args, env2.observation_space.shape[0], env2.action_space, env2)
pg_training1 = agent1.train()
pg_training2 = agent2.train()
for _ in range(600):
result1 = next(pg_training1)
result2 = next(pg_training2)
"""plot"""
plt.plot(range(len(result1)), result1, label='online-policy gradient')
plt.plot(range(len(result2)), result2, label='offline-policy gradient by importance sampling')
#plt.plot(result2)
plt.title('learning curve of pg with {}'.format(env_name))
plt.ylabel('average reward of last 20 episodes')
plt.xlabel('number of episodes playing')
plt.legend()
#plt.grid(True)
plt.savefig('reward-episodes2.png')
plt.close()