[Suggestion] How to make AIs smarter without cheating

from Elemental Forums

In my current game, Magnar sent one of its champions to take me out solo. Ignoring the fact that the champion really had no chance against me, something surprised me about that champion: he carried the scars of over ten injuries.

One advantage that humans have over the computer is that we can take a good look at an enemy army, be it AI or monsters, and have a estimate pretty well whether we can take them on or not with minimal losses. I think we could simulate this in an AI opponent: let it compare its army to a target army/monster and let it run through its circuit-of-a-mind a simulated battle. If it won with acceptable casualties, it goes for the kill, otherwise, it puts off the battle till it raises a greater army/champion.

In this way, I hope that an AI opponent may prove a more able opponent by not needlessly weakening their champions through injuries, nor throwing away armies like canon fodder when it can be avoided.

26,778 views 4 replies
Reply #1 from Elemental Forums Top

That's an excellent suggestion. You get a karma.

Reply #2 from Elemental Forums Top

I already thought the AI DID this kind of stuff, I remember hearing something I think Brad posted, somewhere, saying "The AI is much better at estimating it's odds of success than a human".  If it's not doing what you described, it definitely should be.  Though this does sound like it could get resource intensive by mid game, when there are lots of armies with various enemies to fight.  

Also, I am an idiot for assuming anything, silly me.

 

<edit: fixed typos, I always seem to make a million of them>

WindowBlinds - Free yourself of the default settings, customize your desktop.
Reply #3 from Elemental Forums Top


Yep, I hope and believe the AI is intended to think like this.

Reply #4 from Elemental Forums Top

It's the only think AI can do, in fact - it's called searching the tree of variations. It's how AI plays chess - because each position is a discrete node that can be evaluated with reasonable ease, chess AI with sufficiently strong CPU can nowadays beat the strongest humans.

The problem is the evaluation function, though - in a game like FE, where units can move with relative freedom, each position can lead to many different resulting positions, and the complexity increases exponentially. The computer is always weak in positional analysis though - games like Total War series show this. It does not realize things like "this archer unit is too exposed" or "these units are too vulnerable to cav charge, move them closer to forest".  These observations require long-term, heuristic planning computers are usually incapable to do - and when their tree-searching algorithm starts to see the disasters, it's already too late. 

If you try to "script" the correct behavior, like "move infantry no farther than x from archers" or "charge archers with cavs when they are closer than y, but withdraw if pikemen are closer than z", the resulting behavior will be inflexible and repetetive, and humans will soon learn to exploit it (Total war, again).

Writing AI is hard, because true AI must be adaptive. Sadly, that means having a memory of past successed/failures, pruning the bad behaviors and developing the good, simply having an AI capable of learning. No such thing in current computer games - bar some old quake bots I remember.