def DQN(state_dim,
action_dim,
n_actions,
gamma,
layers=(32, ),
initial_params=None,
target_update_freq=500):
Q = MLPQFunction(state_dim,
n_actions,
layers=layers,
initial_params=initial_params)
Q_target = MLPQFunction(state_dim,
n_actions,
layers=tuple(layers),
initial_params=initial_params)
if initial_params is None:
Q_target._w = Q._w
operator = DQNOperator(state_dim, action_dim, gamma, Q_target,
target_update_freq)
return Q, operator
n_actions = 10
# torch.manual_seed(300)
# np.random.seed(300)
# Create BellmanOperator
operator = MellowBellmanOperator(kappa, tau, xi, gamma, state_dim,
action_dim)
operator2 = mellow(kappa, tau, xi, gamma, state_dim, action_dim)
# Create Q Function
layers = [l1]
if l2 > 0:
layers.append(l2)
Q = MLPQFunction(state_dim, n_actions, layers=layers)
Q2 = mlp(state_dim, n_actions, layers)
Q._w = np.random.randn(Q._w.size)
w = Q._w
w = torch.randn(w.size).numpy()
Q._w = w
Q2._w = w
weights = torch.randn(5, w.shape[0], requires_grad=True)
samples = np.random.randn(10,
1 + state_dim + action_dim + 1 + state_dim + 1)
samples[:, -1] = 0.
samples[:, action_dim + state_dim] = np.random.random_integers(
0, n_actions - 1, size=samples.shape[0])
val = Q.value_weights(samples[:, 1:1 + state_dim + action_dim],
weights.detach().numpy(),