OODA Loop, AI, and Performance

[PatrickAWlson]

And how they might go together.

 

OODA: Observe, Orient, Decide, Act.  Every pilot is taught this.  The AI obviously does something like this.

AI: The AI OODA loop is obviously a tight loop.  Just watch the control surfaces constantly moving at a rate that no human would ever do.  

Performance: it has been stated that AI is the primary consumer of CPU.  The AI is obviously in a tight loop.  No surprise then, right?

 

Proposal:

Break the AI into two parts.  OODA loop and implementation of the decision.

Time box the OODA portion and the decision implementation portion.

 

The OODA loop is performed periodically, with a time gap in between.  Maybe once per several seconds outside of combat and more frequent in combat.  The output of the OODA loop processing is a decision.

 

The rest of the AI processing is the implementation of the decision.  That is where the changes in control input are performed.  These should also be reduced as the AI will be focused on implementing the current decision and not rethinking it - for a little while at least.

 

Results:

The AI processing of the OODA loop will be less frequent and more human like.  Decisions will be seemingly well thought out.  CPU time will be saved by traversing the decision tree less often.

 

The AI processing of the decision implementation will also be less frequent (hopefully).  First, because the decision will change less frequently.  Next, because corrections will be less frequent.

 

Remember that we are talking about a computer here.  A computer will traverse a decision tree in far less than one second.  Decision trees can be pared for the circumstances - attack, evade, formation flying, etc., so all possibilities need not be considered at any time.  The first part of "Observe" will be to pick the right tree.  This is where the AI can change its mind.  By limiting the frequency of this process, you avoid the AI changing its mind multiple times per second.  Less processing.  Less CPU usage.  More believable behavior.

 

Edited February 3, 2022 by PatrickAWlson