Hacker News Reader

Comment by ryandrake | Hacker News Reader

It seems to always get into a rut where one design lucks out and dominates generation after generation, with no mutations producing anything even close to working. Like, the top ten don't change after hundreds of generations. Maybe this is just an attribute of genetic algorithms. They quickly zero in on something kind of good, and then get stuck at this local maxima. Or maybe I need to just play around with the Mutation Rate and Mutation Size settings.

View on HN

>Maybe this is just an attribute of genetic algorithms

It happened to crocodiles and it could happen to you too.

Crabs too https://en.m.wikipedia.org/wiki/Carcinisation

In general, one should keep the mutation rate really low to allow the population slowly change over time. High mutation rate will lead to local optima quickly but also very hard to get out. Low mutation rate will require significant more generations but in general result in better adaptation.

Sounds like you're running with a single elite clone. That's a really bad idea in genetic heuristics for exactly the phenomenon you mention.

Is the correct number 0 or >1?

Note my use of the term "heuristic" and not "algorithm." There is not a correct number. The correct approach is to fiddle with the parameters; record good populations; occasionally restart from scratch... we call this "hyperparameter tuning" to make ourselves feel better about the process.

If you keep going long enough they'll probably turn into crabs.

It seems like more should change than just the shape. I'd wager that a slower car with more power may be less likely to get stuck in ruts. But it seems that the power and speed don't vary, just (barely, after a few generations) the shape.

Edit, I scrolled down and it covers the genome:

• Shape (8 genes, 1 per vertex)

• Wheel size (2 genes, 1 per wheel)

• Wheel position (2 genes, 1 per wheel)

• Wheel density (2 genes, 1 per wheel) darker wheels mean denser wheels

• Chassis density (1 gene) darker body means denser chassis

It basically lands on a two-wheeled medium-bodied shape and doesn't seem to make much progress after that. Power and speed would be interesting variations.

What happens in an evolutionary algorithm depends on what you write it for. This is a fun toy, but what it does specifically is explore a very limited simulation of evolution by natural selection. Metaheuristics aimed at optimization have a lot of techniques aimed at not stalling out on a prematurely converged design, as well as improving other desirable properties of the population, at the expense of any pretense of fidelity to real-world evolution processes.

This is why I said it needed crossover and got downvoted into oblivion :) turns out it at least tries to have crossover, so maybe the genome doesn't translate to crossover doing anything relevant.

I wasn't fooling. Think about it for a second, if your process involves a lot of crossover that means large sections of working genome will be passed on. If the ONLY mechanism for changing anything is mutation, then mostly you're just breaking what works.

That's what you're describing, so I'd look at how the genome is constructed to understand why it's not doing more.

More improvement needs wild environment changes.

Reply on news.ycombinator.com