Machine learning fundamentally involves learning from the data and making conclusions/decisions about a given problem. It utilizes the following popular approaches.
Supervised learning: In supervised learning, a given data set is trained to achieve the desired output using labels, classification, and regression methods. For example, images of different animals, objects can be labeled and passed in algorithms for overall correct identification. The power of supervised learning lies in its ability to scale the available data to predict future outcomes, based on learnings of the sample data.
Unsupervised learning: This approach involves scenarios where the input does not involve labeled data however, a pre-defined model is applied to the data to achieve insights as results. One of the main techniques used in the approach is clustering. The unsupervised machine learning is used mainly for generating insights from the given data-sets and detecting trends and patterns.
Semi-supervised learning: It is a middle path approach of supervised and unsupervised machine learning where the available data-sets include both, labeled and unlabeled data. This method is very useful for avoiding data-labeling exercise for massive data-sets. Furthermore, excessive data-labeling may induce a bias in the model. Including more unlabelled data essentially increases the accuracy of results, at the same time-saving time and costs.
Reinforcement learning: Here, the system is trained in an environment through trial and error approach to take the decision, based on the rewards obtained. The machine “learns” from experiences and uses optimum knowledge to make correct decisions.
Deep Learning: learning like a human brain
Deep learning may utilize supervised, unsupervised, and semi-supervised learning to build the artificial neural networks, which is nothing but an information filtering model similar to the human brain mechanism. Neural networks apply layers of filters which learn from previous layers to form the output which acts an input for the next layer.
Deep Reinforcement Learning
Combining RL and Deep Learning enables solving RL problems more effectively. A single RL agent can solve many complex tasks with deep learning. Deep reinforcement learning utilizes deep learning and reinforcement learning through forming deep learning models with reinforcement learning algorithms to solve complex problems which were formerly unsolved.
Recent research conducted at the University of Technology, Sydney explored a deep reinforcement learning scenario where a model-free RL algorithm was formed with a Bayesian deep model. In this research, deep kernel learning is programmed to learn the complex action-value functions instead of the usual deep learning models. This means leveraging more uncertainty to take advantage of the replay memory. Several experiments have shown that the algorithm proposed in this scenario outperform the deep Q-network. Similar approaches will be explored in the future to use RL for generating dynamic decision making in complex environments which involve large data-sets, agents, and environments. Real world examples where deep Reinforcement Learning can be applied are robotics, healthcare, and trade execution.