ABSTRACT
Deep reinforcement learning (DRL) systems combine the techniques of reinforcement learning (reward maximization) with the artificial neural networks of deep learning (DL). This combination allows a well trained model to operate on much more complex data sets than a typical reinforcement learning system. For example, input data from a camera of sensor stream is high dimensional and would be difficult for a traditional RL system to solve. However, a DRL system can represent the policies for these higher-dimensional choices with its neural network, enabling it to perform much better than RL or DL systems alone. DRL systems have been successfully applied to previously unsolved challenges, such as the game of Go, beating professional poker players in Texas hold’em, and mastering Atari games from pixel inputs.
However, while DRL systems have performed well in the previously listed examples, challenges impede the application of DRL systems on real-world problems. The primary difficult is the infeasibility of allowing the agent to freely and frequently interact with the environment set due to matters of safety, cost, or time constraints. This manifests as the agent having to (1) interact with an inaccurate simulation of the expected environment or (2) interact with a static environment that is a snapshot of a more volatile one – weather conditions or trading markets.
