Hi,
Please someone explain what is the output of the predict method actually representing in the reinforcement learning, precisely in the below code:
def train(self, batch_size = 32):
#training using replay buffer - take a batch of experiences from the buffer and feed them to the neural network
# lets say the batch size is 5 - so we take 5 tuples of values of state, action, reward, new state and done and pass them to the neural network one by one
# after this is done, you take the next action and train the neural network and so on, since we are working in batches, weights are updated after each tuple has been passed in the batch
minibatch = random.sample(self.memory, batch_size)
for experience in minibatch:
state,action,reward, next_state, done = experience
# X, Y : state, reward
if not done : #if game is not yer over:
# use bellman equation to approximate the target value of reward
target = reward + self.gamma*np.amax(self.model.predict(next_state)[0]) #[0]means selecting the inner list
else:
#game over
target = reward
target_f = self.model.predict(state) #what is the output of the neural network for the given state and all the actions - output of predict function is a nested list - size of nested list [batch size, number of actions]
#update the target_f (expected reward)
target_f[0][action] = target #updating the reward for given state and action - approximated using bellman equation
self.model.fit(state,target_f, epochs = 1, verbose =0) #1 epoch because stochiastic gradient descent
if self.epsilon > self.epsilon_min:
self.epsilon *= self.epsilon_decay #as you are gaining experience - rely less on randomness and trust our knowledge more - gradually decreasing this effect