Recursive gametheory as chains of predict and want of scalar vars

  • 0 Replies


  • Bumblebee
  • **
  • 38
Example: I'm trying to build an AI that auto adjusts mouse sensitivity while you move the mouse (like speed based mouse gestures) to get the cursor where you want it earlier. If sensitivity is too high at the wrong times, mouse will go past where you wanted it and you'll keep moving it until it converges to that point, so its uncertain whats too much or too little sensitivity at each fraction of a second.

The person is smarter than AI so will more often cause the cursor to go the speed the person wants it to go, even if AI doesnt understand and sets sensitivity stupidly while it learns. So if we know how fast the person wants to move the physical mouse, just divide to get the correct mouse sensitivity.

But how to know how fast person wants to move the physical mouse? Person would normally move it exactly as fast as they want to move it, so its more of a statement about person wants AI to do something so person didnt have to move the mouse a certain speed. Its recursive gametheory.

If we add another AI to predict physical mouse speed and sensitivity (cursor = mouse*sensitivity), that might tell us something about what the person wants when the predicted mouse movement differs from actual mouse movement, and other combinations.

These combinations got ever more complex, until I thought of "Recursive gametheory as chains of predict and want of scalar vars".

Each mind can predict or want about another mind, and recursively.

Example: person predicts physicalMouse moves at speed .7

Example: AI predicts physicalMouse moves at speed .6

Heres where it gets interesting: A second AI WANTS (person predicts physicalMouse)-(AI predicts physicalMouse) to equal 0. That means the second AI wants the other AI's prediction of physicalMouse to be accurate or the person to move the mouse like that prediction.

Recursing from there, we can chain vars to refer to a mind predicting the accuracy of another mind about a certain var, and each var may be generated by an equation (such as that minus) represented as the prediction of "the mind of the equation". Then minds could PREDICT the output of the equation or WANT it to be a certain value.

I'll need a mind or equation that sums mouse speeds from when mouse starts moving to stops moving, since something should WANT that to be faster considering some function of distance.

Example: X predicts that Y wants Z to predict that Y predicts X is .7

Example: Y predicts X is .8

This might be expanded with digital-signatures for larger scale uses, but mostly it will be used for many millions of vars predicted and wanted per second, a kind of math used with combinations of neuralnets.
« Last Edit: August 19, 2016, 07:32:33 pm by Ben.F.Rayfield »