I was thinking about how to create an AI for an RTS game. Critique it if you will. My idea was to have a cellular automota based brain (or something equivalent) for each type of unit that is run constantly during the game. For example if the game has tanks and grenadiers, there are two types of brain that can be evolved with GA for each type of unit. If you have 100 grenadiers and 10 tanks then you would create instances of the appropriate brains for each unit. The inputs for each brain could include current position, nearest 10 other units, surrounding terrain information etc, and the outputs are of course what the unit will actually do. One problem I can think of is that the GA would take ages to train. What do you guys think?

I think that Regression and AI are not the same thing.

Quote: Original post by apocalypstic
One problem I can think of is that the GA would take ages to train.

Some other problems/questions are:

1) Designers like to have numbers to tune so they can get the desired behavior out of the units.
2) Designers like to have special tuning numbers on specific campaign maps to get certain not normal results.
3) how does unit coordination come in to play here? Where's the strategic planning. i.e. 2 units together might fall into a GA behavior. But sometimes the default behavior is not desired from those units in a skirmish game. The skirmishAI might want to suicide them to achieve a higher goal
4) how does this work with player versus AI controlled units? i.e. there are basic intelligences that player and AI controlled units need to share: auto-targeting, pathfinding. At what level does your GA AI work and how does it allow for player AND AI controlled units to work simultaneously.
5) If you decide at the end of the project that: "you know, having infantry hide behind tanks doesn't really work" how do you get that, and only that, to change?

-me