Novel Computational Intelligence Method of J* Surface Generation for Fast Optimal Decision/Control Design

四十多年来，动态规划方法一直被认为是在非线性、随机决策/控制环境中设计最优过程（策略）的唯一通用方法。在现实世界中，几乎所有的计算机决策和/或控制程序都受益于动态规划方法的应用。不幸的是，对于直接应用于现实世界的问题，其计算需求甚至超过了现有的计算机硬件能力。好消息是，近年来，被称为自适应批评的计算智能方法已经被开发出来，它实现了动态规划的良好近似，从而显着提高了决策和控制策略的质量，这些策略在实践中是可以实现的。这个项目将使计算智能方法更进一步地完成这项任务。PI将尝试在这一过程中注入更多类似人类的智能。背景如下：已知人类能够学习复杂决策/控制任务的最佳解决方案；此外，他们学习的越多，也就是说，他们积累的“经验”越多，那么他们就能够迅速想出一个（近乎）最优的解决方案来解决新遇到的问题，只要它与他们曾经遇到过的问题相似。拟议的项目试图模仿这种能力。特别是，该项目旨在开发一种计算智能方法，根据该环境中现有设计的知识，在假设的问题环境中有效地为附加问题设计最佳控制器。该方法的关键成分将是J*表面发生器（J*SG）。J*SG是一种包含人类“经验”的设备，包含先前遇到的决策/控制问题的最佳策略知识，以及在遇到类似问题时从中推断出（近似）最佳策略的能力。关键的研究任务包括开发将问题规范表示为J*SG的有用方法，以及J*曲面的表示以促进相关（最优）策略的提取。计算智能方法（例如，神经网络，模糊逻辑，遗传算法等）将被用作完成此任务的工具，以及用于实施J*SG。拟议的研究将在波特兰州立大学的西北计算智能实验室（NWCIL）进行。NWCIL将通过为俄勒冈州MESA（数学、工程、科学成就）项目的教师培训做出贡献来扩大规模。该项目旨在增加非裔美国人、西班牙裔、美洲印第安人和女性学生在数学、科学和工程领域的就业机会。

The method known as Dynamic Programming has been known for over four decades to be the only general approach for designing optimal procedures (policies) in non-linear, stochastic decision/control con-texts. Virtually all computer decision making and/or control procedures in real-world problems stand to benefit from application of the Dynamic Programming methodology. Unfortunately, for direct application to real-world problems, its computing requirements would outstrip even present computer hardware capabilities. The good news is that in recent years, computational intelligence methods known as Adaptive Critics have been developed that implement good approximations of Dynamic Programming, and thus significantly enhance the quality of decision and control policies that are achievable in practice. This project will take the computational intelligence approach to this task one step further.The PI will try to inject more human-like intelligence into the process. The context is as follows: humans are known to be able to learn optimal solutions to complex decision/control tasks; further, the more of these they learn, i.e., the more 'experience' they accumulate, then they are able to quickly come up with a (nearly) optimal solution to a newly encountered problem, so long as it is similar to ones they have experience with. The proposed project seeks to emulate this capability. In particular, the project seeks to develop a computational intelligence methodology that efficiently designs optimal controllers for additional problems within an assumed problem context, based on knowledge of existing designs in that context. The key ingredient of this methodology will be a J* Surface Generator (J*SG).The J*SG is the device that is to contain the equivalent of human 'experience', knowledge of optimal policies for previously encountered decision/control problems, and the ability to infer from them an (approximately) optimal policy when a similar problem is encountered. The key research tasks include development of useful ways to represent the problem specification to the J*SG, and the representation of the J* surface to facilitate extraction of the associated (optimal) policy. Computational intelligence methods (e.g., neural networks, Fuzzy logic, genetic algorithms, etc.) are to be used a stools for this accomplishment, as well as for implementation of the J*SG.The venue in which the proposed research is to be performed is the NW Computational Intelligence Laboratory (NWCIL) at Portland State University. The NWCIL will expand by contributing to the training of teachers in the Oregon MESA (Mathematics, Engineering, Science Achievement) program. This program is designed to increase the numbers of African American, Hispanic, American Indian, and women students in careers in the field of mathematics, science, and engineering.