Genetic Algorithms.., A Primer

It is Charles Darwin’s 200th birthday. So I guess there must be loads of posts on GA, strewn across the Internet. Genetic Algorithms are special to me because they are the prime reason why I am doing whatever it is that I am doing.

This post is a light intro to Genetic Algorithms, sans its terms. Genetic Algorithms are delightfully simple. GA is an approximation algorithm, that is, they are used in situations where either the optimal solution is unnecessary or the complexity of the problem does not permit it. Sometimes, the Genetic Algorithm may produce the most optimal result also. Since examples are the best way to learn an algorithm, here’s the example problem we are to solve. Balancing moments

Given to you is a configuration of balances as shown in the figure. The problem is to balance the configuration by adding weights on either side. You are provided with 10 weights measuring 1,2,3……10 kg each. Each weight is to be used exactly once. Thanks to GreatBG for this puzzle. His puzzles are always most entertaining. This can roughly be equated to the traveling salesperson problem. In the problem, the salesperson is to visit all cities (and comeback ), in such a sequence so as to minimize the total distance covered. The most naive solution is to try all possible sequences and see which one gives the minimum cost. For a map of n cities, that comes to n! combinations. A similar approach would work in our case also, but then, when the map or the number of weights is larger, the technique, takes so much more time. A more professional approach would be ruffle up your knowledge of elementary physics and reduce the problem to a set of linear equations and then solve them. Naaa.

Genetic Algorithms

My code, in C++, is here. Here’s the pseudo-code, sourced from wikipedia

1. Choose initial population
2. Evaluate the fitness of each individual in the population
3. Repeat until termination: (time limit or sufficient fitness achieved)
1. Select best-ranking individuals to reproduce
2. Breed new generation through crossover and/or mutation (genetic operations) and give birth to offspring
3. Evaluate the individual fitnesses of the offspring
4. Replace worst ranked part of population with offspring

Initializing the Population

The algorithm maintains a set of solutions. Not optimal solutions, just solutions. For example, in the case of the TSP, it may be a set of random ordering of the cities. In our problem, a set of random configuration of weights. In the program, I maintain the configuration as a string, where the weight alloted to each slot is represented by a character, with ‘A’ being 1, ‘B’ being 2 and so on. The weights are added bottom to top and left to right. One point to be noted here is that all solutions in the set must stick to the constraints of the problem. No city must be re-visited, no weight can be added twice.

Evaluating the Population

This is the key part of the algorithm. Once a way is found to evaluate each member of the ‘population’, the problem is half complete already. In our case, the “fitness” will be based on the resultant moment in the system. Therefore, best solution will have 0 fitness, the resultant moment in the system would be zero.

Until the required precision is obtained, or a maximum limit on the number of iterations is reached, we continue by doing the following opertions.

Crossover

This is where we actually begin to understand why it is called Genetic Algorithms in the first place. According to Darwin’s thery of Natural Selection, the best members of the population mate to produce better offsprings. Similarly, in our algorithms, the best solutions found so far mate with each other to (hopefully) produce a better solution. In code, this translates as the merging of substrings from different members, chosen from among the best solutions found so far. Point to be noted is that if done improperly, this could result in an offspring, that is not even a valid member of the population. It might violate the problems’ constraints. In TSP, and in this case, the same city/weight might get listed again, which cannot be allowed. So we use a special crossover method called, Partially Matched Crossover(PMX in code) which is beyond the scope of a primer.

Mutation

Some members of the population are randomly chosen to be mutated. Mutation is a process by which the offsprings get a trait that is inherited neither from the father nor the mother.While translating into code, this simply means swapping two characters within the string. You can choose more complex mutation techniques if desired. Again, the problem constraints must be satisfied. Mutation is required mainly to increase the variety of solutions in the population. Variety is very important.

Natural Selection and Elitism

Once these operations are done on the parents, a new generation of solutions are born. To make sure that the best solution in this generation is atleast as good as the best one in the previous generation, we introduce “elitism”. Normally, all the individuals in the previous generation are replaced by the members from the newer generation. But by having a non zero elitism value, a small portion of the best solutions from the previous generation survive on in the next generation.

Going by the paradigm of survival of the fittest, we now select the best solutions from the both the generations and these form the actual next generation. (My code is slightly varied. It does not use natural selection, the members of the younger generation simply replace all but the elite members of the elder population)

Local Maxima

Genetic Algorithms suffer from the problem that it might converge to a non-optimal solution. Consider that we are looking for the solution that is at the top of a mountain. It is possible that the population end up on an elevated hill, far away form the mountain. This is possible if the various values are not tuned properly. The algorithm will not get out of this situation by itself. the best way to go in this kind of a scenario is to simply start again.

People, Blogs, Conferences and Applications 