Optimising Constraint Model Selection through Monte Carlo Tree Search and Machine Learning

约束编程的思想是，一个问题可以用一组约束来表示，即几个变量之间的逻辑关系，并且产生问题的解涉及满足所有定义的约束。约束规划已被广泛应用于优化问题、计算机图形学和验证等领域。然而，要将约束编程应用于特定的问题或领域，必须首先将其建模为约束满足或优化问题，这涉及到基于所得到的解必须满足的决策变量集对问题进行建模。在圣安德鲁斯，已经创建了一条管道来简化约束建模和求解的过程，其中可以用一种称为Essence的高级语言来定义问题的抽象规范。该流水线通过一系列的转换将输入的本质规范转换为约束规划模型。本项目旨在研究不同的人工智能技术，如蒙特卡罗树搜索、强化和深度学习，以及它们在本质层上的应用，以便在没有领域特定知识的情况下有效地识别模型并将其转换为其最优等价类。在过去，很少有研究花在改进约束编程的建模方面，这是可以在性能方面获得巨大收益的地方。在此较高级别应用小优化可以在解算器级别实现巨大收益。

Constraint Programming is the idea that a problem can be represented by a set of constraints, a logical relation amongst several variables, and producing a solution to the problem involves satisfaction of all of the defined constraints. Constraint Programming has been applied to a large number of fields including optimization problems, computer graphics and verification. However to apply constraint programming to a particular problem or domain, it must first be modelled as a constraint satisfaction or optimisation problem which involves modelling the problem based upon a set of constraints on decision variables that the resulting solution must satisfy.At St Andrews, a pipeline has been created to simplify the process of constraint modelling and solving where an abstract specification of the problem can be defined in a high level language called Essence. The pipeline then transforms the input Essence specification into a Constraint Programming model through a series of transformations.This project is about researching different AI techniques such as Monte Carlo Tree Search, Reinforcement and Deep Learning and their application at the Essence level to efficiently identify and transform a model into its optimal equivalence class without domain specific knowledge. Little research in the past has been spent on improving the modelling side of Constraint Programming and this is where big gains in performance can be made. Applying small refinements at this higher level can achieve huge gains at the solver level.