Date: 08/22/05 (Algorithms)    Keywords: programming

Hi, my goal improve solution quality (or have the quality found/converged in less iterations) by running the sequential (heuristic agent like) algorithm in parallel in an island/colony mode. Each agent/ant colony searches for solution independently and after a period of time they exchange data with each other. This model is very much like genetic algorithm/programming.

Currently what I have is at S stages , each colony sends the best solution it founds to a central place, this center place then sends back the best solution to everyone and each colony will add a large amount of pheromones to this new solution upon receiving, so basically each new colony will a new pheromone config that will lean toward the new solution. I have implemented many other connection models such as ring (colony A sends to colony B, B->C, C->A etc), and use less greedy method by not sending the best but higher quality solution has a higher probablity of being used etc .

I do achieve better in average solution qualities , for example 3 colonies are better than 2 and 2 colonies better than 1 (sequential). But with 4 , 5 colonies I see the results are worst than 1 colony. Furthermore, the difference is not by much. I hope to get a large differences in average solution quality to show that parallel colony like this is beneficial as it increases the search space and its cooperative sharing manner.

Any idea or suggestion ? Thanks in advance

Source: http://community.livejournal.com/algorithms/64140.html