# Coming to a billboard near you: self-evolving adverts

How mathematicians use evolution to find the best solution to a problem.
16 May 2017

## Interview with

Visar Shabi, Brainlabs

## new-york-1930656_1920.jpg

Maths can be used to explain our decisions, but can it be used to influence them? Well, maths could be beginning to make waves in the world of advertising. Using clever mathematical techniques called genetic algorithms, scientists are able to create self-evolving adverts inspired by Charles Darwin’s theory of evolution. Visar Shabi works on genetic algorithms at Brainlabs and he spoke to Georgia Mills...

Visar - Genetic algorithms are a technique used to solve these really difficult mathematical problems called optimisation problems. Within optimisation problems you usually have a lot of parameters that you need to vary and it’s not immediately obvious how to vary those parameters to get you to your optimal solution or your sweet spot. So, genetic algorithms are really efficient ways of getting to the best solution possible. The naive way of solving them, of course, would just be to try out every different possible set of variables until you find the set which works and genetic algorithms completely skip that and taking a cue from evolution.

Georgia - So how does evolution come into it?

Visar - The way evolution typically works is that within species you have variation, and that variation lends to different members of that species being more successful. The more successful an animal is the more likely it is to pass on its genes, kind of propagating those good genes that allow it to be successful. We take that use that as inspiration of genetic algorithms by using it to come up randomly with solutions for how to essentially pick all the different parameters in our huge optimisation problem. So once we randomly pick different solutions we see which ones are the best and then, by ranking all the best ones, we get to take all the best features from the ones at the top of our ranking and then those get passed on. But by taking all the best bits for all the different solutions that we get, we get hopefully what is the best solution overall.

Georgia - I see. So evolution in the wild, for example, a rabbit might have ten rabbit babies maybe the one that’s the fastest at running, or the biggest will be the one that goes on to survive. It will have ten children, the best of that will survive, it will go on, so on, and so on, and so on until you get the ultimate rabbit I suppose.

Visar - The ultimate rabbit for its environment - yeah exactly.

Georgia - This is what you want to do? You want to take this out of evolution and apply it to other things?

Visar - Exactly, yeah. Apply it to various problems where there is one way to be best suited for some environment.

Georgia - In evolution there’s this natural variation but there’s also DNA mutates. This is when there’s an error in the code and it changes. So do you incorporate this into your model?

Visar - Yeah, exactly right. We do introduce random mutations into our solutions. So once we’ve generated a random solution and that’s turned out to be the best, for successive generations we’ll enter random mutations to kind of try and mix things up a bit. The reason we do this is - the technical term is to kick you out of your local maximum. And a nice way to think about this is if you’re trying to find the peak in a mountain range and you ended up on a smaller mountain top, by introducing mutations we might kick you off of that smaller mountain top and try and get you onto the highest peak.

Georgia - I see. So, for example, if you didn’t do the mutations you might get stuck in this little zone and any deviations will send you downwards; they’ll be worse? But if you introduce this way of more random things happening you can get something totally different which might be even better?

Visar - Absolutely. It’s kind of exploring and seeing if there’s something better out there. Exactly right.

Georgia - OK. So could you give me an example of how this could be used in practice?

Visar - A really cool example of how it was actually used in real life was a coffee company which had an advert up on Oxford Street. They had lots of different images as well as lots of different bits of text and they would combine these to make different adverts. Then they would show this on a bus stop; one of those bus stop ads in Oxford Street. And then by looking and registering how long people would stare at those ads, they would determine if the ad was successful or not.

Once they’ve got enough data on how successful certain ads had been they would take features from the most successful ones and mash them together to try and make even more successful ads, and they would repeat that process.

Georgia - I see. So something like how big the text size is, the colour, the image, the words used; all of that put together and it just sort of keeps going and keeps going until it has the ultimate ad?

Visar - One of the things they found that worked really well were pictures of hearts, but what they did find I think is that the text that it came up with was fairly nonsensical. That may have been the reason for people staring for a lot longer.

Georgia - I was going to ask because I often stare at ads that make me angry because I’m furious with them. Could this not be leading us down the wrong path?

Visar - I think there’s still a lot of refinement to go in the area of autogenerated ads. But, like they say in advertising, if it’s got you talking about it, it’s done the job.

Visar - It will be interesting to see where it goes. Certainly there’s a lot more data now available in order to come up with things like auto-generated ads, but it will interesting to see whether it’s something that actually takes off in the real world.

Georgia - My dad’s a copywriter. You’re going to be putting him out of business here.

Visar - I think he’ll have a job for a long time to come. If you were able to actually see the ads that genetic algorithms produce, but it’s certainly exciting.

Georgia - Are there any other applications this kind of thing could be used for?

Visar - Absolutely. Outside the world of advertising I know Unilever designed one of their nozzles using a genetic algorithm. Then just within other areas of advertising we use it at Brainlabs all the time in order to come up with solutions to big mathematical problems that we’ve got.

Georgia - I guess it made sense copying what nature has been doing for millions of years - it’s what worked pretty well so far?

Visar - Yeah. I think humans are doing OK, so it’s working pretty well.