This is something I've been toying with for some time and struck me as rather interesting. I can't guarantee the technique would be useful for all or even many AI agents, but I still think it's worth discussing. Has anyone ever experimented with using a split-level decision-making entity? A lot of neuroscience research has shown that we, as humans, have a low-level, 'reactionary' intelligence and a higher-level 'planning' intelligence. Approaching the issue of AI this way strikes me as quite beneficial, for, say a twitch-based game, shooter or the like. On one hand, we can spread out some complicated AI over a frame or two without it standing around and being retardedly easy, and on the other we have the ability for agents to *react* in a fairly realistic way-- i.e. a split-second decision that may or may not be the *best* course of action (although one that still makes sense.) One of those 'uncanny valley' deals. I was thinking the reptilian half of the process would essentially be a basic finite state machine, with a setup similar to something you might have found in Quake. Nothing hyper-complicated, just you basic moving, shooting and the like. A planner like that found in F.E.A.R. would then begin to formulate a cunning plan, (and you could theoretically go hog wild on some of this, as you could do it over something like a tenth of a second without it being out-of-place) at which point control would be handed over to the planner for execution. Good idea? Bad idea? Ever tried it? Love to hear back.

A planner-based system that constantly replans based on changing environmental criteria is actually doing just that. If you introduce a stimulus that requires urgent reaction (grenade landing nearby) it invalidates the longer-term plan in favor of the immediate "brain stem" action.

Quote: Original post by InnocuousFox
A planner-based system that constantly replans based on changing environmental criteria is actually doing just that. If you introduce a stimulus that requires urgent reaction (grenade landing nearby) it invalidates the longer-term plan in favor of the immediate "brain stem" action.

I hadn't really thought of it too much from that perspective, but I was thinking of making a model slightly closer to what 'actually' happens per se. While you're kinda referring to general dynamism in decision making and reacting to the environment, I'm interested in actually trying to simulate it at a much coarser level. Is this more redundant that I thought, then?

After the incident that caused the immediate reaction to take place, couldn't you then resume your higher level planning, but obviously with new inputs (like there is a player that knows where you are)?

That's roughly how I go about a deathmatch, making a high level plan of where I'm going to go with what weapons and what I'm gonna do when I get there. Certain things may alter that plan, like a group of friendlies jumping in a vehicle and asking me to join (that would be an almost instant reaction), but my longer term plan may still be there. Maybe these guys will take me closer to where I want to go.

Even reactive short behaviors take some computation and consideration of alternatives (possibly having some mapped out ahead of time).

The grenade comes your way and you have to jump, but which way is usually significant (ie- behind something that will block the shrapnel).

Some actions are reflexes (like bringing up your hand to block a blow or ducking) and can be wired directly to specific short term situational patterns.

Planners can turn into computing pigs because of all the alternatives you have to consider) especially in a rich environment where there are multitudes of things to interact with. Just matching patterns to find something that you have a matching solution (or set) for adds up quickly.

Social interactions increase the complexity exponentially.


The constant reevaluation as situations change makes it a real nightmare as you often have to start all over and recompute again (best laid plans of mice and men...). Figuring out how to determine whether the situation has changed significantly enough to warrant reevaluation is a task in itself.