An Insect-Based Approach to the Dynamic Task Allocation Problem

The thesis aims to develop and implement Ant Task Allocation (TAA), an algorithm based on the model of division of labor observed in ant colonies to solve a scheduling problem.
The performance of ATA are compared with other multi-agent algorithms in the literature. The thesis argues that for several classes of problem instances and for a given objective function, the developed algorithm gets better results than other considered algorithms.
The problem studied is non-deterministic scheduling, with parallel machines and information available only during the resolution of the problem. An industrial environment may be an example of this kind problem: trucks go out of an assembly line and they must be assigned to paint booths. Booths may have the same characteristics or differ in their speed of painting. Any color change implies a time delay and a cost. The problem is complicated by the fact that no information on the color of the trucks is known as long as these do not come off the assembly line. The purpose of the problem is to assign trucks to booths minimizing the makespan, ie the time elapsed from the beginning of the first activity to complete the last task of the system.
The proposed algorithm, which is based on the work of Cicirello et al. and on the model presented by Bonabeau et al., inspired by the method of division of labor observed in insect colonies, defines a set of behavioral rules of the booths.
Each cabin is treated as a single agent (ant) that requires and paints trucks (activity to play). The result is a plastic system for the changes in the experimental environment where agents tend to specialize on a particular type of work according to their characteristics and their status. To support this thesis, the performance of the presented system was compared with other multi-agent algorithms through an empirical analysis of two classes of instances of the problem: a large factory with identical cabins and a medium-sized cabins with different characteristics.
The comparison is made with the ATA, the solution proposed by Cicirello et al., another insect-based algorithm proposed by Campos et al., a market-based algorithm proposed by Morley and a non adaptive algorithm introduced as a reference point for comparing the performances. Particular attention was given to experimental conditions that have seen the use of a generator of instances of the classes of the problem, a search for optimal parameters using genetic algorithms, and rigorous statistical analysis.
For completeness, the experimental analysis shows the contribution that each introduced rule leads to the objective function.