Читать книгу Handbook on Intelligent Healthcare Analytics - Группа авторов - Страница 43

Problems that are goal-oriented can be solved by taking decisions in the order of a sequence is performed by algorithms in reinforcement learning [11]. Here, learning can be carried out based on the reinforcement series, and modeling is based on decision processes such as Markov. These algorithms play a major role in finding solutions for the problems where the decisions to be made need to be in sequence and their applications extend to management of resources and controlling traffic light. Major problem in such algorithms is the preparation of the environment for training the data and performing tasks based on them. When a deep neural network (NN) is combined with reinforcement learning, it is referred to as deep reinforcement learning. The major goal of such algorithms is creation of software agents who learn themselves for establishing the policies in a successful manner. Additionally, they help in attaining the rewards pending in long term. In terms of performance, deep reinforcement learning solves the problems including sequential tasks that are complex such as robotics and computer vision. The need of more training data along with time for training for attaining better performance is a major drawback of deep reinforcement learning as they are computationally expensive. Data can be also be secured with usage of grids [6, 21, 24].