Was musing the other day at the dinner table, thinking of concepts for using GAs for AIs. I was wondering if we could start a discussion here on the possibilities of doing so. The game that I've been thinking of making is a 2D inertia-physics space sim, and using GAs for the AIs to "evolve" into a playing style that is best against the user in the SP campaign seems like a unique idea. My thoughts include the binary genetic code representing things like aggressiveness, favored tactics (jinky maneuvering, setting up good shots, etc.) and teamwork (AIs can send messages to each other, so whether AIs "work as a team" or not would be something to consider). I'm at a loss, though, as to how they should evolve. Genetic modifiers I know, but in terms of where to apply them? All the enemies (should) be killed each level, so it's not like they survive to reproduce. Maybe ones that do the best on each level have a higher likelihood of being carried on and mutated for the next one. If anyone else has thoughts on this subject, feel free to speak up. =)

Well, a pure genetic algorithm implementation would require *many* iterations/generations of enemies.

What you describe could probably be better implemented with a symbolic list of tactics ordered by priority which is then modified by a kind of reinforcment learning. For example, the enemy could try one tactic (chosen randomly but depending on priority or maybe by some predefined rules) and if it successful this tactic's priority is increased otherwise it is decreased. This would actually be able to adapt during a single game round.

IIRC Dark Reign did something similar to this.

Yeah, I think a "true" GA wouldn't work for anything short of Civ or Spore, probably. =)

But adapting enemies, using a variation upon a GA, might be possible. For example, the possibility I'm considering right now has "points" that AIs have allocated towards certain attributes - like aggressiveness, energy saving, etc., all things that could be characterized as attributes of a player. Shifting those points around (take one from aggressiveness, add one to maneuvering capability) could be used as a kind of hack-GA. =)

You may be able to create a model that competes well against a player, but uses such a strange strategy that it reduces immersion/fun. Just something to remember when using machine learning techniques for agent behavior.

I've used a GA implementation in chess, it works quite well. GA in games is fun, but applying it to the stock market is even better. :D

Hah, maybe I should check up on that...

Well, you really don't have to be stuck on using a GA in binary form. You can have GAs that run on real coded values as well. There's really no issue with iterations and such.

If you're coding in attributes, then what you can do is have an elitist GA. This is kind of unorthodox, but you can try this.

Let's say you have 25 enemy planes/space ships or whatever. You can then initialize them heuristically or randomly at first. Then when a ship gets blown up, the others will iterate, perform crossover and reproduce or simply, re-adapt their behavior. In this sense, the fitness function becomes implicit as to survival. So, the longer a ship stays alive or kills more user ships even, the more it gets to reproduce and the more of its attributes will spread through the population. So, the last ship left will most likely have the best combination.

Then what you can do is use the last ship that was blown up in the last wave of attacks to seed the population of the next wave of attacks.

This case, you'll probably have a generational GA where the worst fit member is constantly kicked out, which will be the ship that gets blown up. While the surviving ships reproduce and are replaced by their offspring.

Now, how does that sound. :p

wow - this sounds EXACTLY like the concept for the game I've been working on... cool! Nice to know someone else thinks the same way.

Quote: Original post by WeirdoFu
Let's say you have 25 enemy planes/space ships or whatever. You can then initialize them heuristically or randomly at first. Then when a ship gets blown up, the others will iterate, perform crossover and reproduce or simply, re-adapt their behavior. In this sense, the fitness function becomes implicit as to survival. So, the longer a ship stays alive or kills more user ships even, the more it gets to reproduce and the more of its attributes will spread through the population. So, the last ship left will most likely have the best combination.

[ ... ]

Now, how does that sound. :p

Sounds to me like the enemy that runs away from and avoids all conflicts will reproduce the most.

For something like this to work then successfully killing some humans should be part of, if not entirely, their fitness evaluation. Killing humans being equal to real life darwinian killing for food.

- Rockoon (Joseph Koss)

That's true. Which is why you then throw in a tournament selection for the surviving ones based on a fitness assigned to them. This can be a combination of the health they have left, the number of kills, etc. That way, though you favor the actual acts of evasion, but you also penalize not killing.