No way this will work over the course of a single one player game. It really needs a centralized server and thousands of players, not to mention a better fitness function than lifetime.

When it does work, it takes about 10 levels for interesting behavior to evolve. However that only happens in 1 in 20 games :-( The AI definately needs some work.

That's an excellent idea about a centralized server. that would make things very interesting.

Fitness is not just lifetime, it's fitness = lifetime + (I_Killed_Knight Reward_Factor)

Reward_Factor ~ 4 max lifetime.

There's not too much special about the formula, or the reward factor, those are just what made it work once in a while :-)

He seems to have really missed the point of machine learning by not trying to collect as much data as possible.

I think yes and no. Something like the centralized server is a good idea to learn from lots of players. Did you have something else in mind for collecting data?

But I think machine learing is about selectively ignoring information, cutting to what's important. But of course you do need enough information to start with.

This could be the start of something really really neat, especially if people have time to work on it.

Perhaps I should convert it to Flash so people can play it over the web, and then learn from everyone at once. What other ideas do you see?

I don't think flash would be necessary, though I guess it's potentially easier to collect a lot of data...

What about creating it into an ultimate medieval tower defense? Perhaps a machine learning War Craft II. I don't know.

And he should coevolve knights, removing the need for humans altogether.

An idea I always wanted to play with: competing learning agents, where the players 'train' the characters somehow (perhaps specifying parameters, or playing against their creations, or selectively pitting one character against each other...), who then independently fight and learn over time. Then you could have big melees, and try to evolve the most useful/fun behaviours.

That's a really cool idea.

The formula should evolve dragons which are running away. He gives higher fitness for dragons which live longer, therefore all dragons which attack the player will be the first which are killed. And giving a big reward for killing won't help much. As only one dragon can kill, all other dragons will still profit from keeping out of reach of the player.

Nice idea, but to impress me it should actually work ;-)

Good point. Perhaps there's some game theory in there. The fitness function does need some work.