def execute(args, params, device):
utils.kill_game_processes()
env = main.make_env(args, params)
result_name1, writer1, net1, tgt_net1, selector1, epsilon_tracker1, agent1, exp_source1, buffer1, optimizer1 = main.make_components(args, params, device, env, 0)
result_name2, writer2, net2, tgt_net2, selector2, epsilon_tracker2, agent2, exp_source2, buffer2, optimizer2 = main.make_components(args, params, device, env, 1)
frame = 0
frame_idx1 = 0
frame_idx2 = 0
eval_states1 = None
eval_states2 = None
date_time = datetime.now().strftime("%b%d_%H-%M-%S")
with common.RewardTracker(writer1, params['stop_reward_player1'], net1, date_time + result_name1 + ".dat", 0, env) as reward_tracker1, \
common.RewardTracker(writer2, params['stop_reward_player2'], net2, date_time + result_name2 + ".dat", 1, env) as reward_tracker2:
# fill histories
main.train(args, params, device, buffer1, epsilon_tracker1, frame_idx1, exp_source1, reward_tracker1, selector1, optimizer1, net1, tgt_net1, writer1, eval_states1)
main.train(args, params, device, buffer2, epsilon_tracker2, frame_idx2, exp_source2, reward_tracker2, selector2, optimizer2, net2, tgt_net2, writer2, eval_states2)
while True:
if frame // args.units % 2 == 0:
frame_idx1 += 1
if main.train(args, params, device, buffer1, epsilon_tracker1, frame_idx1, exp_source1, reward_tracker1, selector1, optimizer1, net1, tgt_net1, writer1, eval_states1):
break
else:
frame_idx2 += 1
if main.train(args, params, device, buffer2, epsilon_tracker2, frame_idx2, exp_source2, reward_tracker2, selector2, optimizer2, net2, tgt_net2, writer2, eval_states2):
break
frame += 1
if args.maxFrames > 0 and frame_idx1 > args.maxFrames:
break
def execute(args, params, device):
utils.kill_game_processes()
env = main.make_env(args, params)
result_name1, writer1, net1, tgt_net1, agent1, exp_source1, buffer1, optimizer1 = main.make_components(args, params, device, env, 0)
net2 = ptan.agent.TargetNet(net1)
agent2 = ptan.agent.DQNAgent(lambda x: net1.qvals(x), ptan.actions.ArgmaxActionSelector(), device=device)
frame = 0
frame_idx1 = 0
date_time = datetime.now().strftime("%b%d_%H-%M-%S")
with common.RewardTracker(writer1, params['stop_reward_player1'], net1, date_time + result_name1 + ".dat", 0, env) as reward_tracker1:
# fill history
main.train(params, buffer1, device, frame_idx1, exp_source1, reward_tracker1, optimizer1, net1, tgt_net1, writer1)
while True:
if frame // args.units % 2 == 0:
state, _, _, _ = env.step((1, -1))
action, _ = agent2([state])
state, reward, done, _ = env.step((1, action[0]))
if done:
state = env.reset()
else:
frame_idx1 += 1
if main.train(params, buffer1, device, frame_idx1, exp_source1, reward_tracker1, optimizer1, net1, tgt_net1, writer1):
break
if args.maxFrames > 0 and frame_idx1 > args.maxFrames:
break
frame += 1
if frame % NET_SYNC == 0:
net2.sync()
def execute(args, params, device):
utils.kill_game_processes()
env = main.make_env(args, params)
net1 = dqn_model.RainbowDQN(env.observation_space.shape,
env.action_space.n)
net1.load_state_dict(
torch.load(args.model1, map_location=lambda storage, loc: storage))
agent1 = ptan.agent.DQNAgent(lambda x: net1.qvals(x),
ptan.actions.ArgmaxActionSelector(),
device=torch.device("cpu"))
result_name = "-" + "-rainbow" + "-scenario=" + args.scenario + "-units=" + str(
args.units)
writer1 = SummaryWriter(comment=result_name + "-player0")
env.reset()
total_reward1 = 0.0
counter1 = collections.Counter()
epsilon = 0.02
frame_idx1 = 0
with common.RewardTracker(writer1, 100, net1, "x.dat", 0,
env) as reward_tracker1:
while True:
frame_idx1 += 1
if np.random.random() < epsilon:
action = [env.action_space.sample()]
else:
state, _, _, _ = env.step((0, -1))
action, _ = agent1([state], [None])
counter1[action[0]] += 1
_, reward, done, _ = env.step((0, action[0]))
total_reward1 += reward
if done:
reward_tracker1.reward(total_reward1, frame_idx1)
total_reward1 = 0.0
env.reset()
net1.load_state_dict(
torch.load(args.model1,
map_location=lambda storage, loc: storage))
if args.maxFrames > 0 and frame_idx1 > args.maxFrames:
break
def execute(args, params, device):
utils.kill_game_processes()
env = main.make_env(args, params)
result_name, writer, net, tgt_net, agent, exp_source, buffer, optimizer = main.make_components(
args, params, device, env, 0)
frame_idx = 0
date_time = datetime.now().strftime("%b%d_%H-%M-%S")
with common.RewardTracker(writer, params['stop_reward_player1'], net,
date_time + result_name + ".dat", 0,
env) as reward_tracker:
while True:
frame_idx += 1
if main.train(params, buffer, device, frame_idx, exp_source,
reward_tracker, optimizer, net, tgt_net, writer):
break
if args.maxFrames > 0 and frame_idx > args.maxFrames:
break