Learning Based GA and Application in Packing

于洋,查建中,唐晓君

DOI: https://doi.org/10.3321/j.issn:0254-4164.2001.12.002

2001-01-01

Abstract:Packing problems are categorized as combinatorial optimization problems with strong application background. Except for their intrinsic NP-hard computational complexities, packing problems also involve many constraints and the shape complexity of packing objects and packing space. This paper is concerned with the research on global optimization algorithms based solution for packing problems.One category of packing problems involves global optimization with complex functions of piece by piece smooth continuation. The traditional optimization methods hardly deal with such problems effectively due to the nature of ill-conditioned functions. Recently, genetic algorithms have shown some promising for global optimization, but their efficiency to locate a precise result has not been quite satisfied. This paper, borrowing some idea from natural creatures with capability to learn skills of survival (avoiding dangers and achieving better states) from the environment, has proposed and developed an algorithm, named Learning-based Genetic Algorithm (LGA), to overcome the difficulties the other GA faced. The LGA has made use of the local analytic information of the functions through a Learning Operator to form the Quazi-Newton directions, then let every individual to be able to search for a local optimum efficiently. By comparison with traditional GA, Simulated Annealing Algorithm and Complex Method on typical testing problems, the LGA has shown very good global results in terms of high precision and efficiency. A multi-object optimization model is formulated on simplified satellite cabin packing problem. By comparison on an case of such packing problem constructed with known optimal solution, and another produced with random data and unknown optimal solution, LGA is superior to the Multiplier Algorithm and an Improved GA in term of solution quality and efficiency. In conclusion, the proposed LGA has shown some potential to deal with packing problems with good expectation. Combining a heuristic random searching strategy with local optimal algorithms is effective solution for complex optimization problem.