I have never tried reinforcement learning in my life.

I'm planning to apply it in robotics.

I have some experiences using supervised learning mainly deep learning. So, that's mean I will use neural network in the projects.

We know that supervised learning works on forward propagation to infer and backward propagation to learn.

I have read some articles about reinforcement learning concept . As far as I know, I feel reinforcement learning is doing same as supervised learning, except the robot (the agent) training process in real-time where we (the environment) give reward (for true positive/negative) and punishment (for false positive/negative) from its action (infer).

Unlike supervised deep learning where the splitted dataset for training, validating, and testing have been defined. I mean the X_train, y_train, X_test, y_test have been defined. The reinforcement learning works similiar like that, but the X data wouldn't be defined until the agent gives action and y data wouldn't be defined until the environment gives reward or punishment.

So basically, the agents doing forward propagation, the environment (such as we are as human) giving backward propagation to them whether it's a reward or a punishment based on its action (infer).

We can conclude that, reinforcement learning is basically supervised learning with binary classification problem (reward/punishment) where the X data and y data will be defined in real-time.

The question is, please confirm that my statements were true. So were my statements true? Correct me if I'm wrong.

  • $\begingroup$ No, this is completely wrong, just one difference, in RL there is an environment, in Supervised Learning, there is no environment. $\endgroup$
    – Dr. Snoopy
    Commented Oct 9, 2023 at 6:39
  • $\begingroup$ @Dr.Snoopy In some ofthe broad ideas, it is not completely wrong. Most RL can be split conceptually into the rules for calculating values to associate with states and actions, and rules for learning from those values, and the latter is essentially supervised learning (with some constraints - mainly that the algorithm must be capable of online learning i.e. forgetting older incorrect data) $\endgroup$ Commented Oct 9, 2023 at 6:48
  • $\begingroup$ No, I say completely wrong because the RL setup is wrong , in RL you do not do supervised learning of actions, and there are different RL algorithms from supervised learning (Q Learning, Policy Gradients, Value Iteration, etc). Some of these algorithms do not do supervised learning at all. $\endgroup$
    – Dr. Snoopy
    Commented Oct 9, 2023 at 7:08
  • 2
    $\begingroup$ This seems to be kind of a duplicate of Can supervised learning be recast as reinforcement learning problem? $\endgroup$
    – nbro
    Commented Oct 9, 2023 at 21:54

1 Answer 1


Your statements are mostly incorrect, there are very large differences between reinforcement learning (RL) and supervised learning (SL).

In SL, you have labels that should be the correct answer that a model predicts, in RL you have rewards (which are continuous, not binary), and these rewards do not tell you the right answer.

In RL, you predict actions in an environments based on learning actions on an accumulated reward, in SL there are no actions and there is no environment at all.

While there are many RL algorithms that also use SL, there are RL algorithms that do not use SL at all, like value and policy iteration. This last statement should show you the actual differences between SL and RL.


Not the answer you're looking for? Browse other questions tagged .