def init_v_agent(state_shape, action_shape, action_max, batch_size):
mu_model = Seq_Network([state_shape, 16, 16, action_shape], nn.ReLU(), nn.Tanh())
p_model = Seq_Network([state_shape, 16, 16, action_shape ** 2], nn.ReLU())
v_model = Seq_Network([state_shape, 16, 16, 1], nn.ReLU())
noise = OUNoise(1, threshold=action_max, threshold_min=0.001, threshold_decrease=0.001 * action_max)
agent = NAFAgent(mu_model, p_model, v_model, noise, state_shape, action_shape, action_max, batch_size, gamma=1)
return agent
def init_agent(state_shape, action_shape, u_max, v_max, batch_size):
u_model = init_u_agent(state_shape, action_shape, u_max)
v_model = init_v_agent(state_shape, action_shape, v_max)
v_network = Seq_Network([state_shape, 16, 16, 1], nn.ReLU())
noise = OUNoise(1, threshold=1, threshold_min=0.002, threshold_decrease=0.002)
agent = models.centrilized_naf.CentralizedNafAgent(u_model, v_model, v_network, noise, state_shape, action_shape,
u_max, v_max,
batch_size)
return agent
def init_agent(state_shape, action_shape, u_max, v_max, batch_size):
u_model = init_u_agent(state_shape, action_shape, u_max)
v_model = init_v_agent(state_shape, action_shape, v_max)
v_network = Seq_Network([state_shape, 50, 50, 1], nn.ReLU())
noise = OUNoise(2,
threshold=1,
threshold_min=0.002,
threshold_decrease=0.003)
agent = CentralizedDoubleNafAgent(u_model, v_model, v_network, noise,
state_shape, action_shape, u_max, v_max,
batch_size)
return agent
def init_u_agent(state_shape, action_shape, action_max, batch_size):
mu_model = Seq_Network([state_shape, 32, 32, action_shape], nn.ReLU(),
nn.Tanh())
p_model = Seq_Network([state_shape, 32, 32, action_shape**2], nn.ReLU())
v_model = Seq_Network([state_shape, 32, 32, 1], nn.ReLU())
noise = OUNoise(1,
threshold=1,
threshold_min=0.001,
threshold_decrease=0.003)
agent = DoubleNAFAgent(mu_model,
p_model,
v_model,
noise,
state_shape,
action_shape,
action_max,
batch_size,
gamma=0.999)
return agent
from problems.nonlinear_problem.optimal_agent import OptimalAgent
from utilities.noises import OUNoise
from utilities.sequentialNetwork import Seq_Network
env = NonlinearProblem()
state_shape = 2
action_shape = 1
episodes_n = 250
mu_model = Seq_Network([state_shape, 100, 100, 100, action_shape],
nn.Sigmoid())
p_model = Seq_Network([state_shape, 100, 100, 100, action_shape**2],
nn.Sigmoid())
v_model = Seq_Network([state_shape, 100, 100, 100, 1], nn.Sigmoid())
noise = OUNoise(action_shape,
threshold=1,
threshold_min=0.001,
threshold_decrease=0.004)
batch_size = 200
agent = SimpleNaf(mu_model, v_model, noise, state_shape, action_shape,
batch_size, 1)
def play_and_learn(env):
total_reward = 0
state = env.reset()
done = False
step = 0
while not done:
action = agent.get_action(state)
next_state, reward, done, _ = env.step(action)
total_reward += reward