Hacker News Reader

Comment by flir | Hacker News Reader

I thought genetic programming was going to be The Way Forward for AI, because it seemed to take an intractable problem (write a strong AI) and turn it into one that might just be tractable (write a fitness function that can recognise a strong AI). Then neural networks came from the back of the pack to dominate the field, and here we are. At least we're really good at pattern matching now.

View on HN

Nothing wrong with combining genetic programing with neural networks, there many ways to do it, for example, you can have genetically evolved network that produces actual network(s) (weights, locations, connections), to me biggest is issue is with escaping local minima as efficient/performant as other methods (does not beat deep RL methods), in simple sense it stops evolving and it is really hard to figure out why.

But I don’t really understand. If they could create those virtual creatures in 1994... How come we do not have any real world machines that can do that? Why do our robot lawnmowers still use wheels?

If I had to guess, it's because the physics model that these creatures are being tested against is really simple, and the physics model in your garden is really complex. Also, for this particular application, we shouldn't ignore the fact that wheels are cheaper.

Good point. Though would that same evolutionary algorithm work for adapting to more complex environments?

Thinking of it, why aren’t self driving cars “let loose” in a - real out virtual - playground environment with an ever increasing amount of complexity, given a bunch of rules and goals (“Do not crash into anything”, “Stay within the lane”, “Do not incur a fine”, etc.)... and then left to figure out the “how” themselves?

Sounds like a much more robust approach than teaching them to emulate human drivers (which is what I think Tesla is doing).

Leo... we* don't have any.

Reply on news.ycombinator.com