Optimizing PolyACO Training with GPU-Based Parallelization

A central part of Ant Colony Optimisation (ACO) is the function calculating the quality and cost of solutions, such as the distance of a potential ant route. This cost function is used to deposit an opportune amount of pheromones to achieve an apt convergence, and in an active ACO implementation a significant part of the runtime is spent in this part of the code. In some cases, the cost function accumulates up towards 94 % in its run time making it a performance bottle neck. In this paper we parallelize and move the central parts of the cost function to Graphics Processing Unit (GPU). We further test and measure the performance using the ACO classification approach PolyACO. This GPU based parallelization has a tremendous impact on the performance. The duration of the cost function is reduced to 0.5 % of its original runtime. The over all performance of PolyACO implementation is reduced down towards a remarkable 7% of its original running time - an improvement factor of 14.