Action exploration in D4PG agent.

action_exploration_in_d4pg.py

# Excerpt of D4PG agent object (excerpted).
class AgentD4PG():
    """
    Agent implementing noisy agent
    """
    
    def act(self, states, mode):
    states_v = torch.Tensor(np.array(states, dtype=np.float32)).to(self.device)

    self.actor_local.eval()
    with torch.no_grad():
        mu_v = self.actor_local(states_v)
        actions = mu_v.data.cpu().numpy()
    self.actor_local.train()

    if mode == "test":
        return np.clip(actions, -1, 1)

    elif mode == "train":
        actions += self.epsilon * np.random.normal(size=actions.shape) # use simple normal random noise instead of OU noise
        actions = np.clip(actions, -1, 1)
        return actions