Re: CRT Combat Model (really long)

[Lincoln Peters <peters2000 at mindspring dot com>]

Here's that long list of AI issues I ran into:


1. The computer has difficulty determining the best route from point A
to point B when it has to move away from point B somewhere in the
process.  As a result, it sometimes can't navigate maze terrain at all,
and instead paces along the side of a barrier.  I ran more than a few
games AI-vs.-AI in which one side got so lost that the human (the
self-unit) ran out of supplies and starved to death.  I even ran one
game where both sides (again AI-vs.-AI) got lost and starved to death at
the same time! 
(I guess this problem isn't really part of the AI code, but it must
reside somewhere in the source tree) 

2. All of the items are most useful if carried by another unit.  In
fact, some items are *only* useful if carried by another unit (such as
the Guardian stone).  However, the AI would often see an item, "find"
(capture) it, and then leave it.  This is mostly an issue that the AI
won't pick up inert items (Magic supply kits, Guardian stones, and
Glowing stones), and that does put the AI at a disadvantage. (by the
way, I defined items as "facilities" to try to take advantage of the
advances code, but it didn't seem to help) 

3. I found that in this game, the best defense is usually a good
offense.  By default, the AI would have several units (a garrison)
occupying a place to protect it, but the garrison would never act while
there.  As a result, they would not prevent the place from being
captured; they would just delay it until they were all killed in the
siege.  I sort of fixed it by setting the ai-peace-garrison and
ai-war-garrison both to 0, but that left places wide open to attack when
nothing is standing in the way.  What should really happen is that the
garrison wakes up as soon as an enemy is near and go after it.  In
addition, there should be no garrison if the side does not have the
resources to set one up. 
(should I just increase the damage inflicted by a place so that it can
kill an attacker with one blow?) 

4. The advances code in the AI allows it to effectively handle dwarves
and their ability to build places (colonize), but the colonizing code
seems to override almost everything else.  Oftentimes, the AI will
assign almost every orc hole a "colonizing" plan and build dwarves,
which in turn build more orc holes, which in turn build more dwarves,
etc.  This isn't good practice.  Dwarves take a while to produce, and
there are other humanoids that an orc hole can produce (orcs are equally
effective at combat and take 1/4 the time to produce, elves moves faster
than any other humanoids and can hit harder).  There are also other
places that dwarves can build that are just as useful (nests provide
insects, pentagrams provide imps and pit demons). 

5. The AI doesn't seem to coordinate units to work together very well. 
This problem is most evident when someone is carrying a non-inert stone
(a Lightning stone or Humanity stone).  For example, a dwarf may be
carrying a Humanity stone, and the dwarf intends to build an orc hole,
but the Humanity stone intends to hypnotize an enemy unit.  Unless the
enemy unit is next to the dwarf, the Humanity stone is stuck.  Somehow,
the AI needs to be able to coordinate its plans so that its various
units can do a team effort and not get in each other's way. 
(yes, I'm familiar enough with computer programming to know how daunting
a task that is, but I think that it's important to implement if
possible) 

6. If a unit is capable of normal attacks as well as firing, the AI
doesn't always distinguish the two very well.  I've seen AI-controlled
dragons fire at targets adjacent to them, when they could have done a
normal attack and inflicted twice as much damage (it takes 2 ACP to
attack normally, 4 ACP to fire).  I don't think I've ever seen the AI
fire at non-adjacent targets, either.  That's just not very smart.  It
should fire at targets that are within range, but attack adjacent
targets. 

7. The AI should be smarter about protecting its self-unit.  I tried
running an AI-vs.-AI game earlier today, and then took over the
nearly-defeated side when the AI was ready to give up (all it had left
was the human and a Humanity stone).  As soon as I managed a surprise
attack on an enemy nest, I hypnotized two of the defending spiders,
located the enemy's human, and sent a spider to kill him.  My opponent
had about ten times my point value, and then suddenly lost, all because
the self-unit was left unprotected. (perhaps this wasn't an issue before
because in most games, the self-unit was a capital city that was
protected by a garrison?) 

8. The AI doesn't seem to understand concepts like combat experience
(CXP) or altering terrain.  Usually, these aren't major issues, but in a
big game where the sides are really well-matched, CXP and/or altering
terrain might turn the tide of the game. 

I did make some changes to the game definition that affect defense of
places and garrisons, which I'll send directly to you (Hans) in another
e-mail.