How do you differentiate between backpropagation and reinforcement learning algorithms?

In backpropagation the objective is to minimize the loss, that is to minimize the difference between actual and predicted values. Loss minimization algorithms differ from entropy maximization algorithms. The convergence of the model depends on the minimization of the loss. It is a supervised learning technique most suitable where labelled data is available. We know what we want, have examples of the desired outcome and a distribution that we can model. Backpropagation involves a two step process. A forward pass referred to as the inference and a backward pass based on derivatives. Common use case of backpropagation are text, image and speech applications along with the classic classification and regression tasks.

Contrary to backpropagation, in reinforcement learning algorithms the objective to maximize the reward. The labelled data is not available and the exploration and exploitation paradigm of thinking is used. An agent is exposed to an environment where the agent has to learn through trial and error to take the actions that maximize it's rewards. Deep reinforcement learning uses neural networks for the representation of agents where the agent uses backpropagation to minimize the loss. The loss functions minimizes as the agent learns to maximize it's rewards and minimize it's penalties. For games like chess where the space is exponentially large, DRL is useful as it helps us generalize the agent through neural networks and loss minimization.

The applications that are suitable for backpropagation learning(commonly used in supervised learning in context of ANN) and reinforcement learning are of different nature. Reinforcement learning uses the concepts of finding a policy ( deterministic or stochastic) to be followed and in the process gets rewarded or is penalized depending on the action taken, relearns accordingly to maximize the policy reward. In backpropagation supervised learning, the process is to minimize the error between the actual value obtained and the known target by adjusting the ANN model weights repeated over iterations. Learners or agents in dynamic and uncertain environment require reinforcement learning.

For back propagation algorithms the inputs and output pairs are available, consider an example of a classification or a regression problem with three variables x1,x2 as inputs and y as the output. We have a labelled dataset and the x1 and x2 values are given to the neural network as inputs and the network predicts the y value. The difference in the predicted and the actual y value is then backpropagated through the network using derivatives and the weights are adjusted. In reinforcement learning, the inputs are the environment state and the outputs are the actions the agent performs. The reward/penalty received for the actions is used to update the network. In traditional reinforcement learning a table is kept for input/output pairs.

Backpropagation algorithms are very data intensive. Traditional reinforcement algorithms based on dynamic programming are on the other hand suitable for a fixed sample space with definitive outcomes. Chess is such a game however due to the exponential sample space the traditional algorithms don't work. In cases like these Deep Reinforcement Algorithms are used which encode the agent as a neural network instead of a key-value pair table. To elaborate, in traditional algorithms the agent would make a decision given the state of the environment i.e a particular state of the chess board from a table lookup. In DRL, the state is fed to a neural network and the network decides the outcome. This sample intensive method works wonders. Alpha GO!

Supervised learning problems are suitable for backpropagation algorithms such as image classification and learning language token distributions. On the other hand for unsupervised learning problems based on agent interaction with the environments, reinforcement algorithms are more suitable. It is an interesting point to note that most advanced cases a Deep reinforcement learning solutions is used which encodes an agent as a neural network which minimizes it's loss in order to maximize the reward it receives. The combination of the two approaches allows us to deal with problems where the sample spaces are computationally huge.