def __init__(self,
env,
kind="diff",
gamma=0.8,
weight=0.9,
advantage=False):
if kind == "qv":
qmodel, vmodel = self.create_qv_models(
env.observation_space.shape[0], env.action_space.n)
model = (qmodel, vmodel)
else:
model = self.create_model(env.observation_space.shape[0],
env.action_space.n)
brain = QBrain(model,
kind=kind,
advantage=advantage,
gamma=gamma,
v_selectivity=False,
qnet_soft_update=0.01,
diff_qnet_weight=weight)
brain.compile(Adam(lr=1e-3), ["mse"])
MultiDQNAgent.__init__(self,
env,
brain,
train_sample_size=1000,
train_batch_size=50)
def __init__(self, env, kind="diff", gamma=0.99, diff_qnet_weight=0.7):
model = self.create_model(env.observation_space.shape[0],
env.action_space.n)
brain = QBrain(model,
kind=kind,
gamma=gamma,
v_selectivity=False,
qnet_soft_update=0.01,
diff_qnet_weight=diff_qnet_weight)
brain.compile(Adam(lr=1e-3), ["mse"])
MultiDQNAgent.__init__(self,
env,
brain,
train_sample_size=1000,
train_batch_size=50)
env = TankTargetEnv()
memory = ReplayMemory(100000, v_selectivity=True)
tanks = []
for i in xrange(3):
model = create_model(env.observation_space.shape[-1],
env.action_space.shape[-1])
brain = QBrain(model,
typ="diff",
memory=memory,
soft_update=0.01,
gamma=0.99)
brain.compile(Adam(lr=1e-3), ["mse"])
if i > 0:
brain.transfer(tanks[0].Brain) # make all brains the same initially
tanks.append(TankAgent(env, brain, train_sample_size=1000))
controller = SynchronousMultiAgentController(env,
tanks,
rounds_between_train=10000,
episodes_between_train=1)
taus = [0.01, 0.1, 1.0, 2.0]
ntaus = len(taus)
t = 0
test_policy = BoltzmannQPolicy(0.005)