What are good examples of genetic algorithms/genet

Question

Genetic algorithms (GA) and genetic programming (GP) are interesting areas of research.

I'd like to know about specific problems you have solved using GA/GP and what libraries/frameworks you used if you didn't roll your own.

Questions:

• What problems have you used GA/GP to solve?
• What libraries/frameworks did you use?

I'm looking for first-hand experiences, so please do not answer unless you have that.

Answer 1

There was an competition on codechef.com (great site by the way, monthly programming competitions) where one was supposed to solve an unsolveable sudoku (one should come as close as possible with as few wrong collumns/rows/etc as possible).

What I would do, was to first generate a perfect sudoku and then override the fields, that have been given. From this pretty good basis on I used genetic programming to improve my solution.

I couldn't think of a deterministic approach in this case, because the sudoku was 300x300 and search would've taken too long.

Answer 2

I experimented with GA in my youth. I wrote a simulator in Python that worked as follows.

The genes encoded the weights of a neural network.

The neural network's inputs were "antennae" that detected touches. Higher values meant very close and 0 meant not touching.

The outputs were to two "wheels". If both wheels went forward, the guy went forward. If the wheels were in opposite directions, the guy turned. The strength of the output determined the speed of the wheel turning.

A simple maze was generated. It was really simple--stupid even. There was the start at the bottom of the screen and a goal at the top, with four walls in between. Each wall had a space taken out randomly, so there was always a path.

I started random guys (I thought of them as bugs) at the start. As soon as one guy reached the goal, or a time limit was reached, the fitness was calculated. It was inversely proportional to the distance to the goal at that time.

I then paired them off and "bred" them to create the next generation. The probability of being chosen to be bred was proportional to its fitness. Sometimes this meant that one was bred with itself repeatedly if it had a very high relative fitness.

I thought they would develop a "left wall hugging" behavior, but they always seemed to follow something less optimal. In every experiment, the bugs converged to a spiral pattern. They would spiral outward until they touched a wall to the right. They'd follow that, then when they got to the gap, they'd spiral down (away from the gap) and around. They would make a 270 degree turn to the left, then usually enter the gap. This would get them through a majority of the walls, and often to the goal.

One feature I added was to put in a color vector into the genes to track relatedness between individuals. After a few generations, they'd all be the same color, which tell me I should have a better breeding strategy.

I tried to get them to develop a better strategy. I complicated the neural net--adding a memory and everything. It didn't help. I always saw the same strategy.

I tried various things like having separate gene pools that only recombined after 100 generations. But nothing would push them to a better strategy. Maybe it was impossible.

Another interesting thing is graphing the fitness over time. There were definite patterns, like the maximum fitness going down before it would go up. I have never seen an evolution book talk about that possibility.

Answer 3

I made a little critters that lived in this little world. They had a neural network brain which received some inputs from the world and the output was a vector for movement among other actions. Their brains were the "genes".

The program started with a random population of critters with random brains. The inputs and output neurons were static but what was in between was not.

The environment contained food and dangers. Food increased energy and when you have enough energy, you can mate. The dangers would reduce energy and if energy was 0, they died.

Eventually the creatures evolved to move around the world and find food and avoid the dangers.

I then decided to do a little experiment. I gave the creature brains an output neuron called "mouth" and an input neuron called "ear". Started over and was surprised to find that they evolved to maximize the space and each respective creature would stay in its respective part (food was placed randomly). They learned to cooperate with each other and not get in each others way. There were always the exceptions.

Then i tried something interesting. I dead creatures would become food. Try to guess what happened! Two types of creatures evolved, ones that attacked like in swarms, and ones that were high avoidance.

So what is the lesson here? Communication means cooperation. As soon as you introduce an element where hurting another means you gain something, then cooperation is destroyed.

I wonder how this reflects on the system of free markets and capitalism. I mean, if businesses can hurt their competition and get away with it, then its clear they will do everything in their power to hurt the competition.

Edit:

I wrote it in C++ using no frameworks. Wrote my own neural net and GA code. Eric, thank you for saying it is plausible. People usually don't believe in the powers of GA (although the limitations are obvious) until they played with it. GA is simple but not simplistic.

For the doubters, neural nets have been proven to be able to simulate any function if they have more than one layer. GA is a pretty simple way to navigate a solution space finding local and potentially global minimum. Combine GA with neural nets and you have a pretty good way to find functions that find approximate solutions for generic problems. Because we are using neural nets, then we are optimizing the function for some inputs, not some inputs to a function as others are using GA

Here is the demo code for the survival example: http://www.mempko.com/darcs/neural/demos/eaters/ Build instructions:

• Install darcs, libboost, liballegro, gcc, cmake, make
• darcs clone --lazy http://www.mempko.com/darcs/neural/
• cd neural
• cmake .
• make
• cd demos/eaters
• ./eaters

Answer 4

I made a complete GA framework named "GALAB", to solve many problems:

• locating GSM ANTs (BTS) to decrease overlap & blank locations.
• Resource constraint project scheduling.
• Evolutionary picture creation. (Evopic)
• Travelling salesman problem.
• N-Queen & N-Color problems.
• Knight's tour & Knapsack problems.
• Magic square & Sudoku puzzles.
• string compression, based on Superstring problem.
• 2D Packaging problem.
• Tiny artificial life APP.
• Rubik puzzle.

Answer 5

In a seminar in the school, we develop an application to generate music based in the musical mode. The program was build in Java and the output was a midi file with the song. We using distincts aproachs of GA to generate the music. I think this program can be useful to explore new compositions.

Answer 6

In January 2004, I was contacted by Philips New Display Technologies who were creating the electronics for the first ever commercial e-ink, the Sony Librie, who had only been released in Japan, years before Amazon Kindle and the others hit the market in US an Europe.

The Philips engineers had a major problem. A few months before the product was supposed to hit the market, they were still getting ghosting on the screen when changing pages. The problem was the 200 drivers that were creating the electrostatic field. Each of these drivers had a certain voltage that had to be set right between zero and 1000 mV or something like this. But if you changed one of them, it would change everything.

So optimizing each driver's voltage individually was out of the question. The number of possible combination of values was in billions,and it took about 1 minute for a special camera to evaluate a single combination. The engineers had tried many standard optimization techniques, but nothing would come close.

The head engineer contacted me because I had previously released a Genetic Programming library to the open-source community. He asked if GP/GA's would help and if I could get involved. I did, and for about a month we worked together, me writing and tuning the GA library, on synthetic data, and him integrating it into their system. Then, one weekend they let it run live with the real thing.

The following Monday I got these glowing emails from him and their hardware designer, about how nobody could believe the amazing results the GA found. This was it. Later that year the product hit the market.

I didn't get paid one cent for it, but I got 'bragging' rights. They said from the beginning they were already over budget, so I knew what the deal was before I started working on it. And it's a great story for applications of GAs. :)