Combined optimization algorithm based on spectral clustering and deep dual network

Abstract

The invention belongs to the technical field of computers, and discloses a combined optimization algorithm based on spectral clustering and a deep dual network. According to the combined optimization algorithm, the dimension of the overall behavior space of an intelligent agent is improved through a reinforcement learning deep neural network Dueling DDQN; behavior guidance based on manual optimization combination experience is provided for the intelligent agent by using an undirected graph spectral clustering method; under the condition of frequent change of an optimization target or constraint conditions, efficient calculation speed can still be guaranteed, an optimization result is not affected; the combined optimization efficiency of a road transportation carpooling unit single scheduling link under multiple constraint conditions is improved; a stowage scheduling scheme calculated through an algorithm is checked through actual orders; and checking calculation is carried out by using the same set of freight calculation rules, and the freight is lower than that of the manual level and completely meets all constraint conditions.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a combined optimization algorithm based on spectral clustering and deep dual network. Background technique [0002] Broadly speaking, combinatorial optimization problems are problems that involve finding the "best" object from a finite set of objects, where "best" is measured by a given evaluation function that will Objects are mapped to a certain score or cost, and the goal is to find the object that is worth the lowest cost. Combinatorial optimization problems involving different aspects of various industries in real life usually increase the number of objects in the set due to a small increase in the solution target size or constraints It is very fast, for example, the optimization strategy under K nodes simply relies on exhaustive enumeration with a complexity of O(K!), making the search impractical. [0003] In recent years, with the development of artificial intelligence...

AI Technical Summary

Problems solved by technology

[0002] Broadly speaking, combinatorial optimization problems are those that involve finding the "best" object from a finite set of objects, where "best" is measured by a given evaluation function that will Objects are mapped to a certain score or cost, and the goal is to find the object that is worth the lowest cost. Combinatorial optimization problems involving different aspects of various industries in real life usually increase the number of objects in the set due to a small increase in the solution target size or constraints It is very fast, for example, the optimization strategy under K nodes simply relies on the complexity of exhaustive enumeration to be O(K!), making the search impractical

[0003] In recent years, with the development of artificial intelligence, more and more combinatorial optimization scenarios are trying to introduce reinforcement learning tools and combine deep neural network models to solve them. However, unlike simple man-machine games, the actual business scenario requirements are often not met in advance. Fully quantifying the size of the calculation scale cannot clearly give the boundary of intelligent agent calculations, especially for high-frequency, large-scale optimization tasks (such as vehicle loading and scheduling optimization), which need to be planned within a short period of time A strategy that satisfies all constraints and is relatively high-quality and balances each optimization goal (such as transportation and distribution routes, needs to fully consider the balance between receiving time, vehicle type, and transportation costs)

When simply using the popular reinforcement learning neural network for modeling, because some special situations in the actual business cannot be exhaustively listed in advance, the demand satisfaction degree of directly using the reinforcement learning model to participate in the actual business cannot meet expectations