AAI doesn’t learn. Yet this is its main selling point. Its totally counter intuitive to the end user trying to understand the AI, and it cheapens the many marvels of AAI considerably. As a result the learning capacity of AAI and expectations are vastly inflated in the end users mind, and the capacity of the other AIs to exhibit the same behaviour is greatly diminished by this.
If AAI doesn’t learn then how do you explain this?
Of course AAI adapts based on experience, but this is not through understanding, this is through basic statistical information fiddling. The only part of AAI that could possibly ‘learn’ is the construction process, and only regarding offensive and defensive structures. There is some code that involves detecting important areas of the map, but this is not learning either.
The end user however saw ‘learning AI” and rightly so they expect learning AI to be applied to everything the AI does, because of course, it said learning AI, not learning construction choices.
So if AAI doesnt learn, what does it actually do?
Internally AAI has a set of ‘build tables’, weights, values, all stored together. It uses these to make decisions on what to build, for example it will look at unti A and unit B and if A has greater weighting than B then itll probably build A. It might introduce a little randomness so it doesnt always build A exclusively.
When AAI first runs, all these values are equal. All units are considered equal and just as good as every other unit. A giant hulking war mecha is just as powerful and just as good as the tiny scout with the little laser. So when that hulking war mecha fires its many laser cannons and the scout and blows it up, the war mechas weighting goes up and the puny scouts goes down. This way AAI will now build more warmechs than puny scouts as attackers.
This is not learning though. This is inference via statistics, and it only affects construction choices, and mainly attacker choices. Its all based on unit types killing other unit types and modifying weights accordingly.
It’s not new either?!
This isn’t new. AAI was the first to do it, but it wasnt long before NTai and OTAI did it too. KAI even uses a similar approach for evaluating units, only instead of statistical fiddling, it uses algorithms to pre-generate efficiency values to get immediate results.
Well its not all bad
AAI does have some things going for it aside from learning. AAI has algorithms to detect landmasses, it has customized building placement algorithms, defence placement routines, and much more. These benefits are overlooked and forgotten however due to the great emphasis on the ‘learning AI”
4 replies on “AAI does not learn”
Humans like to have a cloud of mystery around their decision making. They claim that they “understand” stuff whereas machines can only “calculate” stuff. But when you think about it: What is learning other then remembering your expiriences.
I personally have no objections to an AI saying it learns, as long as its an accurate statement. AAIs statistical experience is simply not far reaching enough to suggest the AI as a whole learns. Rather that learning is at best limited to attackers and defenders. But it frustrates me that its marketed as a learning AI when it is not. The other AIs have gotten by without capitalizing on this feature which is not unique to AAI, and it cheapens AAIs other attributes which are far more admirable.
Atleast AAI works and doesn’t crash or lag spring like NTAI. 😉
All known crash bugs have been fixed in NTai and have been fixed for months now.