[Sc2-devel] a better AI: strategy and tactics

[Jason D. S. <js...@an...>]

I assume improving the AI is not currently a priority, but as an 
experienced Melee player it is something I am interested in.


The easiest and most boring way to improve the cyborg AI would be by 
introducing non-combat strategy: specifically non-random ship selection. 
  Each ship generally has around 5 ships that it is particularly strong 
against and 5 ships that it is particularly weak against.  As an 
example, the Druuge is a good ship to pick against the Chmmr, while the 
Spathi generally is not.  Simply adding a few ship 'preferences' here 
could make the cyborg vastly more competitive.  More complicated 
strategy might take into account the damage level of the ship you are 
facing off against, or how a particular choice matches up against other 
ships remaining in the opponents' fleet.

But making this type of change reduces the variety of battles that you 
encounter, so it will also make the Cyborg more boring.  Not good.


Improving the tactics of the cyborg should be a bigger priority.  The 
cyborg is already a spectacular pilot and gunner - facing a cyborg Pkunk 
or Druuge can be an excercise in frustration because of this (again, not 
something that is fun).  But in tactics the cyborg is very weak.  It 
should use more correct tactics against its opponents, and also use more 
varied tactics.  Both will add both to the competitiveness and the 
entertainment value of the AI.

I was discussing this with a friend the other day, and tried to codify 
the different tactics that I use regularly.  I came up with 6 of them. 
In order of increasing sophistication, they are: stand, charge, flee, 
circle, get planet-whip, get planetary orbit.

Looking at the current cyborg code, I see four different tactics: pursue 
(charge?), avoid (flee?), entice (circle?), and wait (stand?).  As I 
recall the AI is fairly incapable of using the planet: it occasionally 
gets planet-whip but never keeps it for long, and never orbits the planet.

To design the AI as an expert system, I would teach the AI to use the 
above six tactics, teach it to recognize the opponent's tactics, and 
then add code to select the current tactic.

A simple way to do this is to choose our tactic based on the opponent's; 
something like
   stand => circle/charge
   charge => flee/get planet whip
   flee => circle/charge/get planet whip
   circle => circle/charge
   planet whip => charge/stand
   orbit planet => charge/circle
but this is probably similar to the current algorithm, and will fail in 
many cases.

Something more extensive, but still logically simplistic would take into 
account the current ship and the opponent's ship as well as the 
opponent's tactics.  In the most brain-dead case this could take the 
form of a 25x25x6 matrix of tactical choices; however this would have a 
lot of redundancy and still wouldn't be as good as a human tactical 
agent.  It would, however, be easy to put in the ship-description 
data-files rather than directly in the code, thus being more modular and 
modifiable.

It is also desirable to change tactics from time to time and/or choose a 
somewhat random element in picking a tactic.  In most situations there 
are two or three reasonable tactics, and if one doesn't work you should 
try another.  This is probably the area where the current cyborg is most 
deficient.

A 25x25x6x3 matrix may sound daunting and unmaintainable, but in the 
grand scheme of things I think it should be manageable.  However, such a 
scheme would still leave many things out, and at that point adding more 
complexity is too much.


Ideally the cyborg could deduce the correct strategy from the basic laws 
of the system and its knowledge of each ship's abilities.  For instance 
it should be able to deduce that a ship that is faster and has longer 
range than its opponent should maintain an ideal distance and then open 
fire.  Coming up with a full list of such rules will take a real expert 
(even this rule is overly simplistic and only works against the 'stand' 
tactic).  There may be a few additional values that the cyborg would 
like to know that could be measured by human experts and then put into 
the ship description files.


For me, the Supox-versus-Yehat battle is an example of how increased 
tactical sophistication can make the game more interesting.  When I 
first started playing (against human opponents), the Yehat would 
generally win.  Soon the flee and circle tactics were developed, and 
things swung over to the Supox's side.  When we discovered the tactic of 
getting planet-whip, things were balanced out again.  However, with a 
combination of intercept-and-flee tactics, the Supox can defeat this. 
With the introduction of the most sophisticated tactic I know, planetary 
orbit, the Yehat was able to again recover an even number of victories.

But could we have known all that just by looking at the ships' 
statistics?  The supox can fire backwards and sideways to its vector of 
acceleration, so it is a good candidate for flee and circle tactics. 
The Yehat has good acceleration and low top speed, so it is a good 
candidate for planetary tactics.

Maybe.

-jason short