Guiding Algorithm Selection and Model Formation with Case-Based Reasoning

摘要

One of the key objectives of AI research is to find ways to solving difficult problems more efficiently and effectively. In general, any type of problem-solving algorithm can also be viewed as a type of search through a space of feasible and infeasible solutions. In the past decade or so, many different types of search algorithms were introduced, algorithms that have been found useful for a wide variety of tasks ranging from engineering to business. These algorithms include Genetic Algorithms (GA) that models a search problem by borrowing concepts from biological genetics. GA is a type of stochastic search that uses genetic operators such as crossovers and mutations. There are also Constraint Satisfaction Problem (CSP) algorithms that model a problem as a set of variables and constraints. CSP uses constraint propagation and domain reduction techniques to perform a quick search for a feasible solution. Simulated Annealing (SA) algorithms are yet other types of problem-solving algorithms. They model a problem using concepts from the physics involved in annealing to search for a near-optimal solution. There are also Mathematical Programming, Tabu Search and many others. Besides the problem of algorithm selection, each type of search algorithm also has many different types of variations and tweaks for fine-tuning, such as addition heuristics to guide the search. Furthermore, each type of problem-solving algorithm will require a different problem model before they can be applied. For example, if GA is used, the problem must first be modeled as chromosome-like structures. However, selecting the "right" problem-solving algorithm and producing a "good" model is not at all easy; even for skilled and experienced AI practitioners. Very often, a trial-and-error approach is used to discover the best combination of algorithms and models to match a particular problem at hand.