def select_action(self, state):
state = transform(state).unsqueeze(0).to(device).float()
with torch.no_grad():
alpha, beta = self.net(state)[0]
action = alpha / (alpha + beta)
action = action.squeeze().cpu().numpy()
return action
def forward(self, inputs, action = None, z_out = False, xesn_out=False, future_out=False):
if future_out:
return self.future(inputs)
if z_out:
inputs = transform(inputs).unsqueeze(0).to(device).double()
B = inputs.shape[0]
_, _, _, _, zs = self.model.vae(inputs)
zs = zs.view(B, -1)
return zs
if xesn_out:
inputs = transform(inputs).unsqueeze(0).to(device).double()
action = torch.from_numpy(action).unsqueeze(0).double().to(device)
x_esn_input, x_esn = self.model(inputs, action)
B = inputs.shape[0]
S = torch.cat((x_esn_input, x_esn, torch.ones((B, 1)).double().to(device) ), dim=1)
return S
def expand(self, curNode: TreeNode):
if (curNode.game.getReward()[1] == True):
return -1
policyInput = transform(curNode.game)
policyInput = np.expand_dims(policyInput, axis=0)
(intuitionProbs, intuitionValue, _) = self.intuitionPolicy(policyInput)
intuitionProbs = intuitionProbs.numpy()
intuitionProbs = np.squeeze(intuitionProbs)
validMoves = curNode.game.getValidMoves()
intuitionProbs = self.sanitizeActionProbs(intuitionProbs, validMoves)
curNode.intuitionProbs = intuitionProbs
for move in validMoves:
nextState = curNode.game.getNextState(move)
nextNode = TreeNode(nextState)
curNode.edges[move] = TreeEdge(
nextNode, move[0], move[1],
intuitionProbs[self._flattenMove(move)], 0.)
return intuitionValue.numpy()
if __name__ == "__main__":
agent = Agent()
agent.load_param()
env = gym.make("CarRacing-v0")
env = Env(env)
running_score = []
state = env.reset()
for i_ep in range(100):
score = 0
state = env.reset()
hidden = [torch.zeros(1, RSIZE).to(device).double() for _ in range(2)]
for t in range(1000):
state = transform(state).unsqueeze(0).to(device).double()
action, hidden = agent.select_action(state, hidden)
state, reward, done, die = env.step(action *
np.array([2., 1., 1.]) +
np.array([-1., 0., 0.]))
if args.render:
env.render()
score += reward
if done:
break
running_score.append(score)
logger.info('Ep {}\tScore: {:.2f}\t'.format(i_ep, score))
logger.info('Avg Score: {} + {}'.format(np.mean(running_score),
np.std(running_score)))