Genetic Algorithms - Jeremy Fisher
Vložit
- čas přidán 30. 06. 2024
- This talk is part of Cerner's Tech Talk series. Check us out at engineering.cerner.com/ and @CernerEng
Genetic Algorithms: Programming by the Seat of Your Genes!
The term Genetic Algorithms sounds intimidating to most, a subject obviously beyond the comprehension of anyone with fewer than two advanced degrees. But in truth, genetic algorithms are - like the biological evolution that inspired them - little more sophisticated than trial and error, and their power to solve problems with complex constraints makes them a tool worth having. This talk will bring genetic algorithms out of academic papers and expensive textbooks and teach those of us in industry what's needed to put them to use.
About the Speaker
Jeremy Fisher is Director of Advanced Engineering and Data Science in DST's Applied Analytics Group, where he leads a team of data hackers and algorithm junkies. Prior to joining DST, Mr. Fisher was a Group Technical Director at VML advancing brands like Gatorade and Revlon, and before that was Director of Software Engineering at Adknowledge, where he was responsible for the advertiser technology platform. His specialties are fast-paced engineering, internet-scale architectures, and leading the best and brightest engineers and scientists. - Věda a technologie
![Genetic Algorithms In Trading: How To Automatically Generate Profitable Strategies! [FREE TRIAL]](/img/n.gif)
Great talk. It felt weird when there was just silence after any of the jokes.
From 4:58 to 5:40
I was looking for a simple explanation on GA. That was it. Simple
Most other explanations I've found over complicate things and make it feel like the foundation is more difficult than what it is
For the question at 37:00 - The local optima problem is mostly addressed in GA with the mutation. The best mutation methods and factors help in "jumping out" of local optima.
best talk/ tutorial i've encountered on the subject. I was stuck on the subject of encoding until this video.
What I am often missing in disscussions about genetic algorithms is the link to biology. Because what GAs do is harnessing darwinistic and evolutionary principles (obviously). I think there is a lot more from biology that can be applied to GAs, for example how the competition between solutions actually affects the selection and how heavier mutation or larger populations affect the generated solutions.
Also I feel like neural networks and genetic algorithms are the perfect fit for AI. One could use a genetic algorithm to evolve a neural network or have a neural network optimize a genetic algorithm by learning what mutations make the solutions score better. So basically a learning genetic algorithm.
Langkopf Kopf : Yes, like how we are now able to alter our own evolution to some degree.
what do you mean?
"how the competition between solutions actually affects the selection"
I think this is just a fitness function. It is very hard to mimick natural selection in nature because there are very many moving parts, as well as there are no directions in evolution. So they use a fitness function, which is kind of an 'artificial selection'.
"how heavier mutation or larger populations affect the generated solutions."
I think there are entire set of papers devoted in genetic algorithm optimizations. I never read them before, but I can only assume that they optimise various parameters such as mutation or crossover.
possibly.. so the question may be: what is the fitness function in the real world? and since everybody is basically trying to figure that out all of the time, it is definitely not a simple question. But that might show, that the genetic algorithm will only ever find solutions as good as the fitness function is, eg as close as the fitness function is to your real world application. And as that can be incredibly complex, it may be a better start to simulate the real world environment as accurately as possible (real world environment must not necessarily mean the physical outside world in all of its depth and space, but can also be a simple space as a factory or a virtual space like a computer network) and then just let the simulation be the fitness function.
That is exactly what NEAT does. Neural Evolution of Augmented Typologies.
Wow. This was really great. I like those nice visuals and color representations!
So far the GA and neural methods don't really mimick so much. One missing component is a morpholagicl neural network.
Not only do the weights change but the connections too.
The neurons should be input^input
And connections (input!/input).
Also on the out put methods remove any if statements and recode such that the neurons have full control of the out put.
This feels like a practice in front of the speaker's cat instead of the actual presentation.
just wonder, has anyone tried to do something like the traveling salesman problem, with a genetic algorithm without crossover, in my head it would make more sense
you mean just mutation?
You need crossover to exploit your population. If you dont crossover you can just aswell mutate one solution until you get it right. With crossover you get to combine the best solutions and converge a lot faster.
@@georgechristoforou991 no, it still need a selection algorithm to, the problem is in the salesman problem is are the city something that would be like a gene, or would it be more like a basepair, my view is it like a base pair, and the list of cities would be a single gene
@@distrologic2925 bacterias work without crossover, I think you could do a form of crossover, but the problem in the salesman problem is what is really the gene, and I think the whole list would be a single gene, but typically a single place would be looked at as a gene.
@@TheHpsh czcams.com/video/q6fPk0--eHY/video.html or look for "traveling salesman problem, four algorithms" there is also something called Simulated Annealing being tested there, im sure you will find it interesting.
Wtf is wrong with the dude at 44:00 waving the microphone around while talking?? I can't hear SHIT!