I'm trying to write an AI for a tactics game in the vein of Final Fantasy Tactics or Vandal Hearts. I can't change the game rules in any way, only upgrade the AI. I have experience programming AI for classic board games (basically minimax and its variants), but I think the branching factor is too great for the approach to be reasonable here. I'll describe the game and some current AI flaws that I'd like to fix. I'd like to hear ideas for applicable techniques. I'm a decent enough programmer, so I only need the ideas, not an implementation (though that's always appreciated). I'd rather not expend effort chasing (too many) dead ends, so although speculation and brainstorming are good and probably helpful, I'd prefer to hear from somebody with actual experience solving this kind of problem.
For those who know it, the game is the land battle mini-game in Sid Meier's Pirates! (2004) and you can skim/skip the next two paragraphs. For those who don't, here's briefly how it works. The battle is turn-based and takes place on a 16x16 grid. There are three terrain types: clear (no hindrance), forest (hinders movement, ranged attacks, and sight), and rock (impassible, but does not hinder attacks or sight). The map is randomly generated with roughly equal amounts of each type of terrain. Because there are many rock and forest tiles, movement is typically very cramped. This is tactically important. The terrain is not flat; higher terrain gives minor bonuses. The terrain is known to both sides. The player is always the attacker and the AI is always the defender, so it's perfectly valid for the AI to set up a defensive position and just wait. The player wins by killing all defenders or by getting a unit to the city gates (a tile on the other side of the map).
There are very few units on each side, usually 4-8. Because of this, it's crucial not to take damage without gaining some advantage from it. Units can take multiple actions per turn. All units on one side move before any units on the other side. Order of execution is important, and interleaving of actions between units is often useful. Units have melee and ranged attacks. Melee attacks vary widely in strength; ranged attacks have the same strength but vary in range.
The main challenges I face are these:
Lots of useful move combinations start with a "useless" move that gains no immediate advantage, or even loses advantage, in order to set up a powerful flank attack in the future. And, since the player units are stronger and have longer range, the AI pretty much always has to take some losses before they can start to gain kills. The AI must be able to look ahead to distinguish between sacrificial actions that provide a future benefit and those that don't.
Because the terrain is so cramped, most of the tactics come down to achieving good positioning with multiple units that work together to defend an area. For instance, two defenders can often dominate a narrow pass by positioning themselves so an enemy unit attempting to pass must expose itself to a flank attack. But one defender in the same pass would be useless, and three units can defend a slightly larger pass. Etc. The AI should be able to figure out where the player must go to reach the city gates and how to best position its few units to cover the approaches, shifting, splitting, or combining them appropriately as the player moves.
Because flank attacks are extremely deadly (and engineering flank attacks is key to the player strategy), the AI should be competent at moving its units so that they cover each other's flanks unless the sacrifice of a unit would give a substantial benefit. They should also be able to force flank attacks on players, for instance by threatening a unit from two different directions such that responding to one threat exposes the flank to the other.
The AI should attack if possible, but sometimes there are no good ways to approach the player's position. In that case, the AI should be able to recognize this and set up a defensive position of its own. But the AI shouldn't be vulnerable to a trivial exploit where the player repeatedly opens and closes a hole in his defense and shoots at the AI as it approaches and retreats. That is, the AI should ideally be able to recognize that the player is capable of establishing a solid defense of an area, even if the defense is not currently in place. (I suppose if a good unit allocation algorithm existed, as needed for the second bullet point, the AI could run it on the player units to see where they could defend.)
Because it's important to choose a good order of action and interleave actions between units, it's not as simple as just finding the best move for each unit in turn.
All of these can be accomplished with a minimax search in theory, but the search space is too large, so specialized techniques are needed. I thought about techniques such as influence mapping, but I don't see how to use the technique to great effect. I thought about assigning goals to the units. This can help them work together in some limited way, and the problem of "how do I accomplish this goal?" is easier to solve than "how do I win this battle?", but assigning good goals is a hard problem in itself, because it requires knowing whether the goal is achievable and whether it's a good use of resources.
So, does anyone have specific ideas for techniques that can help cleverize this AI?
Update: I found a related question on Stackoverflow: https://stackoverflow.com/questions/3133273/ai-for-a-final-fantasy-tactics-like-game The selected answer gives a decent approach to choosing between alternative actions, but it doesn't seem to have much ability to look into the future and discern beneficial sacrifices from wasteful ones. It also focuses on a single unit at a time and it's not clear how it could be extended to support cooperation between units in defending or attacking.