Thoughts on design of AI

I have a few thoughts all you fledgling AI modders probably might want to discuss.

The current way AI is programmed in SEV is to explicitly define ship design types, adjust percentages of what goes in a ship, how many ships are defining/attacking, how many colony ships to build, etc.

In a detailed game such as SEV the creation of a decent AI thus is very daunting and extremely prone to simple exploits by the player, because it can barely adapt. But how to do it better?

In my humble opinion, AI should be designed in a goal-oriented way, using solely cost-functions, which self-adapt. Now, what does that mean?

The most important thing is a cost-function. For any given game-state it's possible to compute the value of a specific ship design, a specific planet, a specific tech, a sector, just about everything. The funny thing is - all these values can be made comparable!

Currently, cost of techs is only used to compare techs to choose the next one. The cost for a ship design is only used to compare ship designs, etc. However, what if you make it such that the cost of everything compares with each other? Suddenly you are able to decide that instead of trying to build even more research centers to get that nice new tech it's perhap much more important to hold that one sector (containing the only warp point to your home planet)

However, this value computation must go entirely automatically, not by manipulating cost parameters or even functions. This *is* possible, because we have all the numbers there: We have ranges, damage types and values, miss-rates and all that stuff. The only thing missing is a weight ratio that allows to tell what is more important. For example, two otherwise identical cannons - where one fires at double range but does only half damage should not receive the same score. In fact, their value is dynamic. If the enemy has lots of short range ships, the second cannon is more valuable than the first one. If the enemy has faster ships, then contrary the first is more valuable. But how to arrive at these weight ratios?

There are several ways: Build simply several vaiants, send them to battle and see what survives, adjust weights, rinse and repeat. That would represent a very slow learning, hampered by being constantly behind because technology research would constantly add new weapons for which no weights were acquired yet.

Another way is to run a series of combats with randomized ship layouts and in the end calculate which weapons had the most effect, then safe the acquired weights (and conditions under which which weight setup to take) in the scripts and combine it with the above variant.

Now, what about goals? Actually, it's all about cost-functions as well. Define a few goals, such as tech-rush, spamming, defensive, whatever. And then assign values for each goal based on current gaming situation. If the other empires are technologically behind, switch to spamming them (to further hamper them). If you are losing more and more planets, switch to defensive. What each goal actually means, is again a hierarchically nested subset of smaller configuration changes, such as more/less mineral miners, etc.

And all such should go along with a few safety rules. Such as never build more facilities than you can sustain or always have at least one colony ship.

Well, all that are just a few thoughts. I guess we'll never see anything like that in SEV, but perhaps we could at least keep that stuff in mind, when designing our AIs.