Lets create a driving virtual currency, and we have multiple bots, each of them pricing the various functions of a car. Braking, accelerating, steering left or right. They price he virtual currencies against the sensory inputs. But we apply a trick. When the human brakes, the braking bot gets more currency and the other functions get less. So the sensory input are queue up for the braker bot. Brakesd are applied and the ratio in the second difference goes up, the bot is going to make bigger bets that the sensory inputs are worth more than before. Well over a few trials, the sensory inputs associated with the braking function will appear at the from of the queue as opposed to the non-relevant sensory inputs. It is like a neural net, the non relevant sensory inputs loose more often than not and go broke. When they go broke they are re-attached to another bot down stream.
The key is to determine the total amount of bandwidth for the various driving functions, then allocate that to each of the banker bots as the action happens. The bots will match the N number of optimum sensory inputs to the bandwidth, thus optimally allocating the total bandwidth for driving a car. Then the various bots, connect up to the bandwidth bot, will be making bets that deliver bandwidth when not needed, and extract it when needed. All done by queue management. The hierarchy of bots will optimally group the sensory elements to the specified [precision. The clue here is combinatorics, the second derivative is simply the number of combinations that are available. Once the cominations are distributed, the group theory is done.
This works everywhere, like maximally allocating face features so some finite collection of faces are optimally separated for identification.
No comments:
Post a Comment