Quoting zlefin, reply 34
good things to hear sounds like good progress.
On AI; my offer to help write improved ai still stands; and the biggest way to easily make significant improvements in the ai is through genetic algorithms and data mining of the games played by all the players; that allows for much better races and units quite readily.
I appreciate the offer to help.
But unless you have ever actually written a computer game AI in a shipping game I don't think you (or anyone in a similar place) is in a position to understand what it takes to write good game AI.
I've been familiar with expert systems, neural nets, and genetic algorithms for over two decades now. I might be wrong but I'm not familiar with anyone who has been writing computer game AI as long as I have (maybe someone reading this could check). The OS/2 version of Galactic Civilizations used to save bits of data to disk that when the player played future games the AI could gather what type of player he or she was. I even tried to implement a primitive expert system (took it out for performance reasons).
The Windows version of Galactic Civilizations I would record build order, tech tree research order, etc. that could then be used by the AI to improve on its research. Galactic Civilizations II added the ability to run, in a thread, simulated turns (substituting computer AI for the human) to get a guess on what is likely to happen an then "react" to that future even in the present.
And while those things are useful, they are extremely time consuming to write and you can do all that only to discover that none of it matters if the player has figured out they can simply win the game by buying 20 salted porks for their mounted units and kite to victory. Or aim straight to archers, build a large force of them (to get the larger maps) and just massacre every army they play. Or combine spamming henchmen with buying quests and build an uber army.
Game design is always a much bigger effect on how effective computer players are than any sort of theory craft on AI development. It's all about bang for the buck.
Your interest in AI development is commendable. I think you would be better served putting that interest to use in your own game so that you can see what works and what doesn't when it comes into contact with human beings.
Two decades of knowledge is great; it doesnt' mean you're so smart and knowledgeable noone can help improve it. And it doesnt' an educated computer scientist can't have a good understanding of ai. Especially when it's demonstrable that most commercial game ai's can be improved with relatively modest amounts of effort (though i don't know how long it would take for FE).
I know it's quite complicated; but that doesn't mean one can't make improvements through others. Just because certain builds are ridiculously powerful doesn't mean the ai can't be better at fighting those who aren't using the most brutal builds available.
And I well know and understand how game design affects the viability of computer players; and how some of the changes in the civ games have ahd significant effects on how well computer players can do. I've been studying STRATEGY for over 20 years, and i know a darn lot about it.
If I had a game of my own to work on I would, but I don't; I have ideas for games, as many people do, but that's far from having a game to work on for ai purposes, there's a lot of other things in the game that's needed before it gets to the point where you can focus on ai. This leads to one of the obnoxious catch-22's; you dont' have experience so you can't get a job, you don't have a job so you can't get experience.
Neural nets do take too long to do the learning, that's why they wouldn't be great for this; but genetic algorithms are very viable if you do them right, nobody in the business does. At this point I have to think you're just being defensive rather than effective, sorry to say, but it's true. I've mentioned ways that can improve the ai, and are not hard to program, and you refuse them, if they are unclear I can clarify, but they are there. The main thing in the industry that people don't do is COMBINE the results of all the millions of games played frmo all the players into the ai. A learning ai that only learns from each player will of course take up too much processing time and not improve much; but systems which focus on simple hill-climbing algorithms utiliizing the millions of games played to find higher spots on the hill can readily make some improvement. You already have systems wherein the ai reuses races/units to get better at facing you. The simple solution to better ai unit building is to have the ai not just reuse the units you built against it, but to take trends and commonalites and best units from ALL players who play the game and connect to the internet, and combine them; it won't have perfect accuracy, but it's a huge amount of data, and as with most things, there are diminishing returns; so the returns on the first basic processing of that huge amoutn of data are quite high.
I also know about separability, and the degree to which some parts of the game are separable from others. Even if separability isn't perfect, it's often easy to find subsystems in a game which can be improved without the complexities of looking at the entire game. For example the unit designs in FE; while they somewhat depend on what you're facing, you could make some improvements quite feasibly even without such.