Agent AI

[DigitalLibrarian]

We might end up integrating a scripting engine, but before doing that we are going to experiment with mixable behaviors parameterized via the agent class.

There will be a set of agent behaviors that correspond to different parts of an agents ai routine. It will include high level things like:

• Is undead? Undead mindlessly attack anything not undead
• Is dead? Dead men don't walk, but might haunt as a ghost...
• Item preferences
• Psychological dispositions: cheerful, cowardice, desires, dislikes, etc

It also might be beneficial to include a built-in enum type in the parameter set representing the stock AI strategy that should be used for greater variety and control.

We want as much of the behavior to be dynamic as possible. So things like picking the attack moves and agent commands are relegated to a very configurable behavior engine.

Picking an attack will ultimately take into account situational awareness (move factory discovery invoked via agent behavior), and agent class configuration.

Moves should be able to be compared based on configurable heuristics given by agent class, but that isn't as important as the rest and can be handled by a single set of hard-coded heuristics for now.

[DigitalLibrarian]

I imagine the AI move discovery system being flexible enough to allow for a chess-style experiment, where the combatants can look ahead by min max recursion, to see which combat sequences might result in the quickest win. If done correctly, then this move prediction pattern could lend itself to conversational submission as well.

[DigitalLibrarian]

With a rich enough content model, we don't need a scripting language AI.

This ticket remains as a place to design out the ai config surface and mechanics.