Subgoal Identifications in Reinforcement Learning: A Survey

Designing an algorithm to build a program to solve all forms of problems is an attractive idea. The programmers don’t need to spend efforts on figuring out an optimal way to solve the problem, the algorithm itself will explore the problem and automatically finds the solution to the problem. This amazing feature of automatic programming is what makes reinforcement learning so appealing, and have the potential to apply on virtually every single task of our world. Reinforcement learning is a general framework to find an optimal solution for the given task. Its generalization reduces the effort of programmers on mapping a specific task into a reinforcement learning problem, but this feature is also the main performance bottleneck of reinforcement learning. Since there are not many constraints within the framework, the search space of the policy is exponentially proportional to the dimension of the state space. When mapping a high dimensional problem into a reinforcement learning problem, the conventional reinforcement learning algorithm becomes infeasible. Most of the real world problems are high-dimensional, and it is the major limitation for reinforcement learning. Therefore, how to find a way to reduce the search space and improve the search efficiency is the most important challenge. On dealing with a high dimensionality problem, there are two common approaches to improve the performance. One is to reduce the dimensionality; the other is to find a better optimization algorithm. The first approach has drawn much attention in recent years, and many interesting works have been proposed this way to boost the performance of reinforcement learning. This article is going to review some recent advances in the dimensionality reduction approach. Approaches for dimensionality reduction can be classified into two categories. One is value function approximation, and the other is abstraction. The first category approximate the utility function as a specific form of function, usually linear function, and the optimization of value function becomes much more efficient when dealing with linear functions. This approximation not only reduces the cost of optimization, but also provides the abilities to dealing with continuous-variable problems and generalizing the original policy to similar problems. The second category identifies the structure of the problem, represent the problem with higher-level concepts, and remove irrelevant parameters. Approaches within this category identify the sub-spaces of the state space that do not need to include other subspaces when solving the problem. For example, on solving a “make-coffee” problem, we can divide the problem into two sub-problems “reach the coffee maker” and “cook coffee with