pg.py

def training_loop(self):
        #remember our gradients! 
        gradients = np.vstack(self.gradients)
        rewards = np.vstack(self.rewards)
        rewards = self.discount_rewards(rewards)
        rewards = rewards / np.std(rewards - np.mean(rewards))
        gradients *= rewards
        X = np.squeeze(np.vstack([self.states]))
        Y = self.probs + self.learning_rate * np.squeeze(np.vstack([gradients]))
        #update our model after a full-episode, did we win or lose? 
        #Use that as the reward signal
        self.model.train_on_batch(X, Y)
        self.states, self.probs, self.gradients, self.rewards = [], [], [], []
        
if __name__ == "__main__":
    env = gym.make("Pong-v0")
    agent = PGAgent()
    while True:
        env.render()
        action, prob = agent.act(x)
        state, reward, done, info = env.step(action)
        score += reward
        agent.remember(x, action, prob, reward)
        agent.train()