GARs

GARS: An Improved Genetic Algorithm with GARS: An Improved Genetic Algorithm with Reserve Selection for Global OptimizationReserve Selection for Global Optimization

Reviewed by

Paskorn [email protected]://dssg.cs.umb.edu

Yan Chen*, Junglu Hu*, Kotaro Hirasawa*, Songnian Yu**

GECCO 2007

*Waseda University, Japan**Shanghai University

mailto:[email protected]

http://dssg.cs.umb.edu/

http://dssg.cs.umb.edu/

December 16, 08 DSSG Group Meeting 2/19

Outline• Research Issue• Research Approaches• Genetic Algorithm with Reserve Selection

– Premature Convergence– Reserve Selection

• Simulation Results• Conclusion


Research Issues• In applying GA to solve large-scale and complex

problems, GAs often result in an unsatisfactory compromise.– lack of accuracy

• Solutions do not converge to global optimal points.– Solutions may go to the local optimal points – Premature convergence

» The population is trapped in such a local optimal that most of the genetic operators can no longer produce offspring that outperforms their parents.


Research Approaches• There are several mechanisms to overcome

premature convergence.– The increase of the population size

• increases the algorithms running time – The increase of the mutation rate

• What is the mutation rate should be?• In small population with high mutation rate scenario, the

mutated offspring may be eliminated by their parents.– The use of more disruptive crossover operators

• How to perform crossover to avoid premature convergence may be different in different problems.


Premature Convergence

• When premature convergence happens, all the individuals in a population tend to be identical with almost the same fitness value.– The individuals with higher fitness value are selected

in the mating pool more than once, whereas the many less-fit individuals are rejected.


Tournament Selection

Mating Pool

C1

Randomly select n individuals

Cn

The winner is placed in the mating pool

Population


• When tournament selection (size 5) is used to select individuals to a mating pool, – the loss of genetic information is measured by the loss rate (Rloss) calculated

using the ratio of the number of unselected individuals in each generation (Nu) to the population size (N)

• Only about one-third of the individuals are selected to produce offspring, while all the others are died out due to their low fitness. The genetic information has not been fully utilized (only small amount of solutions are used ; diversity is reduced), it is disadvantage to find global optimization.

• There may exists some redundancy in offspring, since they derive from only a small part of their ancestors.


GARS: Genetic Algorithm with Reserve Selection

• Approach– Some of the less-fit individuals should be reserved in

order to increases the diversity in the population– The redundancy offspring should be removed


Reserve Selection

• GARS proposed a new selection called reserve selestion.

• GARS based on the technique called population segmentation, which divides the offspring population into two parts– Non-reserved area– Reserved area


• Non-reserved area (NRA): similar to the population of standard GAs. The offspring in NRA originate from more-fit individuals since they are produced via tournament selection.

• Reserved area (RA): this part maintain a diversified search to explore the global optimum. The size of RA is called reserve size. The offspring in RA is created from the individuals those not having been selected to parents in NRA (low fitness value individuals)

– Selected table is used to label the individuals selected in producing NRARA helps to increases the diversity in the population.


Evolution of Reserved Area• NRA:

– Tournament selection uses crowding distance of the individuals to diversify the population.

• The individual that has higher crowding distance is the winner.


Experimental Results

• Population size =10• The max generation number = 100• The reserve size = 4• The results are the average outcome of 1000

independent runs.



Results

• Given the same population size and generation number, the precision of the proposed algorithm (1.69% and 0.40%) is much higher than that of the conventional GA (14.14% and 7.79%).

• The proposed method takes the advantage of the reserve selection to prevent GAs from premature convergence such that the better solution can always be searched for.


Traveling Salesman Problem

• The traveling salesman problem (TSP)– The problem is given a number of cities and the costs

of travelling from any city to any other city, what is the least-cost round-trip route that visits each city exactly once and then returns to the starting city


• Population size =100• The max generation number = 100• The reserve size = 30• The results are the average outcome of 100 independent

runs.• The testing are all picked from TSPLIB, a library of

110 sample instances of traveling salesman problem

http://tinyurl.com/6l76e9


TSP Results

• The proposed algorithm performs better than the conventional one with a promotion in result fitness by 11.92%, 13.96%, and 15.71% respectively


• The result from kroA200.• In the earlier period of evolution, the proposed method converges more

slowly than the conventional one till the 23rd generation.• In he long run, the proposed algorithm progressively demonstrates its

efficiency, while the conventional method loses its convergence speed resulting from lack of population diversity.


Conclusions• Premature convergence is one of the major problems

faced by genetic algorithms.• The population diversity gets reduced when the less-fit

individuals are losing along with the evolution.• GARS applies reserve selection mechanism to maintain

the diversity in population.• The experiments show that GARS outperforms the

conventional Gas.Future work:

- how different reserve sizes affect the performance- theory analysis the proposed method will be applied to solve wider range of complex problems