def main(unused_argv):
agent = MoveToBeacon8Actions(state_size=_SCREEN,
dqn="prio",
learning=False)
evaluater = Evaluation(50)
try:
with sc2_env.SC2Env(
map_name="MoveToBeacon",
players=[sc2_env.Agent(sc2_env.Race.terran)],
agent_interface_format=features.AgentInterfaceFormat(
feature_dimensions=features.Dimensions(screen=_SCREEN,
minimap=_MINIMAP),
use_feature_units=True),
step_mul=16,
game_steps_per_episode=0,
visualize=_VISUALIZE) as env:
# run_loop.run_loop([agent], env, max_episodes=1)
episodes = 0
while episodes <= _EPISODES:
episodes += 1
agent.setup(env.observation_spec(), env.action_spec())
timesteps = env.reset()
agent.reset()
while True:
step_actions = [agent.step(timesteps[0])]
if timesteps[0].last():
# agent.dqn.decrease_epsilon_factor()
# this is C
if episodes % 5 == 0:
agent.dqn.update_target()
# agent.dqn.replay(32)
evaluater.moving_avg(
timesteps[0].observation.score_cumulative[0],
agent.dqn.epsilon)
break
timesteps = env.step(step_actions)
agent.dqn.save_weights()
# agent.ql.save_q_table()
# agent.ql.print_q()
except KeyboardInterrupt:
print("Exception")
pass
def __init__(self,
state_size,
action_size,
path="Learning/Weights/weights.h5",
new_weights=True,
memory_size=100000,
replay_start_size=6000,
epsilon=1,
epsilon_min=.05,
max_step_for_epsilon_decay=125000*3,
prioritized_replay=False,
alpha=0.6,
beta=0.4,
beta_inc=0.0000005):
self.state_size = state_size
self.action_size = action_size
self.path = path
self.use_prio_buffer = prioritized_replay
if not prioritized_replay:
self.memory = deque(maxlen=memory_size)
else:
self.prio_memory = PrioritizedReplayBuffer(memory_size, alpha)
self.beta = beta
self.beta_inc = beta_inc
# self.beta_schedule = LinearSchedule(max_step_for_epsilon_decay,
# 1,
# 0.4)
self.gamma = 0.95 # discount rate
self.epsilon = epsilon # exploration rate
self.epsilon_min = epsilon_min
self.epsilon_decay = 0.995
self.max_step_for_lin_epsilon_decay = max_step_for_epsilon_decay
self.epsilon_decay_linear = self.epsilon / self.max_step_for_lin_epsilon_decay
self.learning_rate = 0.00025
self.replay_start_size = replay_start_size
self.model = self._build_model()
self.target_model = clone_model(self.model) #self._build_model()
self.target_model.compile(optimizer='sgd', loss='mse')
self.step = 0
if not new_weights:
self.model.load_weights(path)
self.update_target()
self.callback = Evaluation.create_tensorboard()