What are the primary uses for deep reinforcement learning in practical settings, and what are its obstacles?

Applications:

Deep Reinforcement learning, or DRL. It can now learn from unprocessed sensors or photos as input more effectively, allowing for end-to-end learning and expanding the field of applications to include robotics, computer vision, gaming, natural language processing, and more. Numerous fields, including gaming, robots, Natural Language Processing (NLP), transportation, industrial applications, communication and networking, and more, have made extensive use of deep reinforcement learning. Other methods than DRL are typically used in applications; for example, supervised learning and reinforcement learning are used to train AlphaGo.

Games : Since a long time ago, researchers have been using games as testbeds for their ongoing studies of RL algorithms. The research on card games, video games, and board games is covered in this section. There have been notable advancements in AI-powered gaming that lead to DRL. A neural network called TD-Gammon was developed in 1994 to estimate the evaluation function for Backgammon using TD learning. In the 1990s, Buro also authored articles that suggested ways to do Othello programs (accessed on January 11, 2021).

In his thesis, Mannen used experiments to train neural networks to serve as chess program assessment functions. Block et al. extended KinghtCap's learning technique with a wider database by using reinforcement learning in chess engines. Later, Lai demonstrated Giraffe, a chess engine that beat other engines at the time by utilizing Deep Reinforcement Learning with autonomous feature extraction.

Video games : Computer games are the most widely used testbeds for DRL algorithms, and video games are excellent resources that provide video frames as inputs to RL/AI agents. CNNs are used in DRL for video games, much like in the MDP method, to extract elements from the video frames (environment) that the agents can identify. DRL has recently been used by Sega, Nintendo, Xbox, and other companies. Shao et al. conducted a thorough analysis of DRL in video games.

The Arcade Learning Environment (ALE), a classic platform, is a trailblazer assessment platform in reinforcement learning research that offers an interface to several Atari 2600 games. In the 3D environment of Doom, a traditional first-person shooter (FPS) game, players see perspective 2D projections from their places as pixel matrices. The primary difficulty is the partial observability of game states.The primary difficulty is the partial observability of game states. An API is available on the Doom-based platform ViZDoom.

Steering, braking, and acceleration are all movements in the racing game Open Racing Car Simulator (TORCS). It has used as a test platform for games with visual input. Because of its realistic feature, it is also a crucial platform for testing approaches related to autonomous driving. In the well-known sandbox video game Minecraft, users may freely construct buildings and earthworks in a variety of game modes while exploring a blocky, procedurally-generated 3D environment. It offers a wide range of situations, from cooperation and problem-solving exercises to navigation and survival, providing a rich setting for role-playing. 

A customized version of Quake III Arena, DeepMind Lab is a 3D customisable first-person gaming environment for agent-based AI research. It is often applied to investigate how large-scale, partially observable, visually varied landscapes provide challenging challenges for RL agents. In DeepMind Lab, several researchers have developed and evaluated RL tasks and algorithms.

Numerous fields, including research, industry, healthcare, education, and entertainment, employ robotics extensively. Numerous difficult issues, including as perceptrons, control systems, operating systems, etc., are involved in robotics. In robotics, the advancement of deep reinforcement learning leads the way. Numerous applications discussed in the section on gaming, such as AlphaGo, are also related to robotics.

Due to the obstacles of high dimensionality, intermittent contact dynamics, and under-actuation in dynamic object manipulation, dexterous manipulation is one of the most difficult control issues in robotics. Despite this, dexterous manipulation is broadly applicable across a variety of disciplines. Because of this, robotics is discussing it a lot. Nguyen and La reviewed DRL for Robot Manipulation, including its difficulties, current issues, and potential future study areas.

Challenges :

The fact that RL needs a large amount of data to learn from is one of its primary problems. RL agents must interact with the environment and try out various behaviors in order to choose the best course of action, in contrast to supervised learning, where the data is tagged and curated. This may be highly expensive and time-consuming, particularly in settings that are complicated and dynamic.

Ensuring the security and morality of the agents and their activities is another difficulty for RL. The reward mechanism may include flaws or unexpected effects that RL agents can learn to take advantage of, which might be contrary to the intentions or ideals of the human creators or users.