Which is Better Deep Learning or Reinforcement Learning?

August 16, 2022

Which is Better Deep Learning or Reinforcement Learning?

Supervised Deep Learning Vs Reinforced Learning

Artificial Intelligence & Machine Learning

Supervised Deep Learning

Supervised learning solves known problems. Moreover, supervised learning enables algorithms to “learn” from historical/training data and apply it to unknown inputs to obtain the correct output. This is a type of deep learning method that uses labeled data with the aim of training algorithms, classifying data, or predicting outcomes. Learning algorithms are taught how to generalize from training data, then machine learning learns the “how” to do this, and then machine learning can perform unseen situations. Upon completion of the training process, a subset of the testing data is utilized in model testing (Sharma).

The speed at which the model learns is dependent on the datasets holding both the inputs and accurate outputs. It is beneficial in that it allows one to use data output from experience. However, it has a large computation time. It is used in various applications such as speech recognition and object detection. This type of learning maps the function from input to output as shown in the equation below:

y=f (x) [1]

There are various supervised deep learning algorithms namely, artificial, recurrent, and convolutional neural networks

Artificial Neural Network:

This algorithm works similarly to how the nervous system of humans does. It trains on the dataset and learns by performing a comparison between its output to the already known actual value. It aims at decreasing error through weight and bias modification of the interconnection. Error is considered to be the difference between the algorithm’s output and the actual value (Sharma).

Recurrent Neural Network:

This algorithm feeds the output from the initial step as input to the current step. It is majorly used for sequential data. Node connections in the neural network establish a directed graph along a sequence. This sequence is however temporary, which is an aspect provided by this algorithm.

Convolutional Neural Network:

This algorithm consists of numerous layers for processing and extracting meaningful image features (Wani et al. 3). It first finds the image features and uses the relu activation function to break linearity. It then reduces dimensions, computation, and overfitting through pooling. The pooling output is then located into a one-dimension matrix through flattening. Lastly feeds the flattening output into the neural network aiding in further image classification and recognition.

Reinforced Learning

The central idea of reinforced learning is to let the intelligence learn in the environment. Each action will correspond to its own reward, and the intelligence learns by analyzing the data, how to do what in different situations. Reinforced learning is a machine learning method in which the agent acquires a reward for every right behavior and a penalty for every wrong behavior (Alharin et al. 2). Furthermore, the agent learns to perceive and interpret the environment at hand, take actions, and make use of trial and error methods to learn from new situations. The agent becomes programmed to look for both maximum and long-term general rewards to obtain the expected solution.

The long-term goals allow the agent to focus on bigger goals and eventually, the agent continually seeks rewards instead of penalties. Artificial intelligence has adopted reinforced learning in directing unsupervised deep learning in the form of rewards and penalties. Furthermore, it is beneficial in that it allows performance maximization. However, it is not preferable in solving simple problems. It is applied in various applications such as robot motion control and traffic light control, there are various reinforced algorithms as discussed below:

Q-learning:

Moreover, this algorithm does not provide the agent with a policy hence the agent becomes more self-directed in exploring its environment (Alharin et al. 4).

State-action-reward-state-action:

This algorithm provides the agent with a policy that notifies it with the odds of particular actions that might result in rewards.

Deep Q-Networks:

Lastly, this algorithm makes use of both reinforced learning techniques and neural networks. Self-directed in exploring their environment, a concept acquired from reinforced learning. As a result, random samples of previous beneficial actions learned by neural networks form the basis of future actions.

Written by Yuchen Wang