Giving a Bomberman AI intelligent bomb placement

Question

I'm trying to implement an AI algorithm for Bomberman. Currently I have a working but not very smart rudimentary implementation (the current AI is overzealous in placing bombs).

This is the first AI I've ever tried implementing and I'm a bit stuck. The more sophisticated algorithms I have in mind (the ones that I expect to make better decisions) are too convoluted to be good solutions.

What general tips do you have for implementing a Bomberman AI? Are there radically different approaches for making the bot either more defensive or offensive?

---

Edit: Current algorithm

My current algorithm goes something like this (pseudo-code):

1) Try to place a bomb and then find a cell that is safe from all the bombs, including the one that you just placed. To find that cell, iterate over the four directions; if you can find any safe divergent cell and reach it in time (eg. if the direction is up or down, look for a cell that is found to the left or right of this path), then it's safe to place a bomb and move in that direction.

2) If you can't find and safe divergent cells, try NOT placing a bomb and look again. This time you'll only need to look for a safe cell in only one direction (you don't have to diverge from it).

3) If you still can't find a safe cell, don't do anything.

for $(direction) in (up, down, left, right):
    place bomb at current location
    if (can find and reach divergent safe cell in current $(direction)):
        bomb = true
        move = $(direction)
        return

for $(direction) in (up, down, left, right):
    do not place bomb at current location
    if (any safe cell in the current $(direction)):
        bomb = false
        move = $(direction)
        return

else:
    bomb = false
    move = stay_put

This algorithm makes the bot very trigger-happy (it'll place bombs very frequently). It doesn't kill itself, but it does have a habit of making itself vulnerable by going into dead ends where it can be blocked and killed by the other players.

Do you have any suggestions on how I might improve this algorithm? Or maybe I should try something completely different?

One of the problems with this algorithm is that it tends to leave the bot with very few (frequently just one) safe cells on which it can stand. This is because the bot leaves a trail of bombs behind it, as long as it doesn't kill itself.

However, leaving a trail of bombs behind leaves few places where you can hide. If one of the other players or bots decide to place a bomb somewhere near you, it often happens that you have no place to hide and you die.

I need a better way to decide when to place bombs.

Answer 1

The problem you're facing is that your AI never stops to make an intelligent decision about where it should place its next bomb, which leaves it just dropping bombs whenever it can and then working out "shit, shit, what do I do now!?"

Pausing to think

Right now, your AI just wanders to nowhere in particular. Sometimes, however, it should actually be moving to a target. For instance, if it sees a powerup and thinks it can reach it before the player, perhaps it ought to find a safe path to that tile and move there, keeping an eye out for bombs and avoiding them along the way.

The same moving-to-target behaviour can be used when it thinks about where to place its next bomb.

When your AI can place a bomb, instead of just doing so immediately, it should briefly use a search algorithm to choose from the available bomb placement spots based on criteria such as:

• Is it nearby? (so that it doesn't walk to the diagonal opposite corner of the map each time it wants to place a bomb)
• Is there a place I can hide from the explosion - and can I reach there before the bomb explodes?
• Will the explosion blow up tiles and give me powerups?
• Will it potentially attack the player? (an easier AI might avoid this in the early game, an aggressive AI will pursue this - whilst not forgetting about powerups)

When this decision has been made, the AI has chosen a bomb placement spot, and a place to hide from its explosion. It can now walk to the spot, place its bomb, then walk to its hiding place. Once it reaches its hiding place, it might want to keep walking and finding bomb locations (if it has multiple bombs), provided it keeps in mind to keep out of the way of the bomb it walked here to avoid.

On walking, and finding a safe path

You can develop for your AI a single walk-to-point method and use this each time you want it to walk someplace. This method could use the A* search algorithm to find its path.

In order to keep the AI safe, you may want to make it recheck its path each time a new bomb is placed. To avoid explosions, it could check the time til the bomb explodes, and consider a potential explosion tile safe to walk over if the bomb won't explode whilst it's walking through that tile - and if it's not safe, treat it as an impassable tile.

To give your AI faults (so it can actually be blown up sometimes) it should make calculation errors: forgetting about a bomb absent-mindedly, underestimating the explosion size or the time til the bomb explodes, etc.

Side note: You can also make its random walking seem more purposeful by randomly picking spots to walk to, or always having a new bomb placement spot in mind to walk to, instead of just randomly picking an adjacent tile to wander to. That way it won't go wandering back and forth on the same spot as if it can't make up its mind.