Introduction
Deep reinforcement learning has become a flourishing subfield of machine learning in the past decade. Two remarkable and well-known successful cases of using deep reinforcement learning to solve sophisticated games, Atari 2600 and AlphaGo, substantially catalyze the research interest in this direction globally. An important goal of machine learning research is to create intelligent systems that are able to automate complex decision making and achieve human-level control. Deep reinforcement learning holds the promise to be an important component of this system.
Reinforcement learning is an experience-driven autonomous learning method. The essence of it is learning from interaction with the environment. Broadly, reinforcement learning has its early roots in the behaviorist psychology (trial-and-error learning) and optimal control (its solutions using value functions and dynamic programming). These two subfields provide the foundations for the modern reinforcement learning.
While reinforcement learning has made important progresses across various domains, a bottleneck is lack of scalability. Specifically, to successfully implement reinforcement learning for solving real-world tasks, the learning agents confront a challenge of deriving efficient representation of the environment (Mnih et al., 2015). In other words, reinforcement learning per se lacks scalability and is inherently limited to low-dimensional problems, a.k.a. “the curse of dimensionality”.
The advances in deep learning provide a powerful tool to tackle this bottleneck confronted by reinforcement learning. Deep learning has powerful function approximation and automatic feature learning properties, which enables the reinforcement learning agent to effectively handle the unstructured environment. Incorporating deep learning into the reinforcement learning framework gives rise to the so-called deep reinforcement learning. Combining deep representation learning with the reinforcement learning framework makes it possible to learn complex policies in high dimensional environments and solve such complex high-dimensional problems end-to-end.
In this study, we aim to provide a brief overview of the core ideas and algorithms in reinforcement learning and deep reinforcement learning, and summarize recent applications of deep reinforcement learning in stock market research.
The rest of this study is organized as follows. Section 2 introduces the background, core concepts, mathematical set-up, and key algorithms reinforcement learning. Based on this, section 3 discusses the rise of deep reinforcement learning and surveys important and influential deep reinforcement learning algorithms. In section 4, we briefly summarize recent applications of deep reinforcement learning algorithms on stock market research. Then, we highlight primary challenges confronting deep reinforcement learning and discuss promising research directions going forward. Then, we summarize in section 5.
The Rise of Deep Reinforcement Learning and Applications in Stock Market Research
