Topic: News about AI

As is usually the way, once I get myself committed to helping with something life gets in the way, I will be as good as offline until November at least so cannot test much at all.

I have run some AI tests, using normal against very weak, same tribe. There is virtually no difference between them, on games that ended without deadlock there is a 50% of either side winning. Testing done on a variety of maps from small to large on several releases up to but not including the last release.

Top

Quote

Training is going wonderfully now, I often get several major improvements in just one generation. Just a couple of remarks about the AI behaviour:

• The problem that AI builds lots of metalworkshops is probably related to the last change in helpersite algorithm. I´ve managed to train AI to the point where it builds not 40 but "only" about 10-20, but that´s still too much, as it usually can´t supply them all. It also comes with the catch that having not many metalworkshops generally means that the AI builds too few wells. It shows this kind of irrational behaviour only for barbarian metalworkshops. AI also likes to build helpersites but no using site, for example several gamekeepers but no hunter nearby.

• May I suggest considering some major changes to the road building algorithm. Stupidly built roads often increase transport times by more than 10 times, and they tend to block space-consuming buildings like farms. Also, AI is prone to getting nearly deadlocked with stuffed-up bottleneck roads and doesn´t build a second road even if there is lots of space and it would shorten transport times significantly.

• Barbarian AI often builds a battle arena long before it even has a brewery. Also, it builds its first training camp and battle arena simultaneously, which often means that each site gets 2 gold and neither is finished. Some minor changes to the prohibited_till flags might help. Incidentally – what unit are those flags measured in?

Top

Quote

Nordfriese wrote:

Training is going wonderfully now, I often get several major improvements in just one generation. Just a couple of remarks about the AI behaviour:

• The problem that AI builds lots of metalworkshops is probably related to the last change in helpersite algorithm. I´ve managed to train AI to the point where it builds not 40 but "only" about 10-20, but that´s still too much, as it usually can´t supply them all. It also comes with the catch that having not many metalworkshops generally means that the AI builds too few wells. It shows this kind of irrational behaviour only for barbarian metalworkshops. AI also likes to build helpersites but no using site, for example several gamekeepers but no hunter nearby.

To me it looks like a dead branch of evolution - few days ago you mentioned some good mutation in regard to metalworkshops happened but obviously the effects are not that good. But I will do some changes to code and hopefully with further training things will settle down... The problem is the fact that metalworks is upgradeable and it is treated differently....

• May I suggest considering some major changes to the road building algorithm. Stupidly built roads often increase transport times by more than 10 times, and they tend to block space-consuming buildings like farms. Also, AI is prone to getting nearly deadlocked with stuffed-up bottleneck roads and doesn´t build a second road even if there is lots of space and it would shorten transport times significantly.

Yes, I know about this issue. But - it is complicated... Road algorithm is not controlled by genetic algorithm... AI tries to build a road from a flag filled with material, but sometimes it is not possible.

• Barbarian AI often builds a battle arena long before it even has a brewery. Also, it builds its first training camp and battle arena simultaneously, which often means that each site gets 2 gold and neither is finished. Some minor changes to the prohibited_till flags might help. Incidentally – what unit are those flags measured in?

These flags are in seconds. And feel free to modify them. When the training will be finished (or your spare time spend) you can propose new branch with new wai numbers and changes to aihints there as well...

Top

Quote

I just had a look at the values for the prohibited_till/forced_after flags. They´re a mess. They appear to be mostly arbitrary; many dependencies between buildings are not considered at all. These flags seem to be the reason why empire AIs are particularly deadlock-prone. The barbarian quarry is prohibited_till=240 and forced_after=0, which is not the most logical combination.

I´ve made a branch which contains the latest training results and my suggestions for better flags. For now, I only adjusted ports, militarysites and trainingsites; my suggestions for productionsite ai-hints will be pushed tomorrow. I´m revising them completely.

Of course, the new hints will need to be tested before merging them into trunk. I expect there will be some trial-and-error before I find a set of values that solves all the problems of the current flags without creating new difficulties.

Top

Quote

So I am testing wai files from your branch, with code from current trunk, and nothing strange here in regard to metalworkshops - after 3 hours they have 2 or 3 metalworkshops. Is the problem map-specific?

Top

Quote