So we take SARA and put a handful of different kinds of sensors on her. Get all of that data streaming successfully back to a mega-database and get her ready to fly. I can now use the manual control system and fly her around our test ground (the laboratory parking lot at night) and 'instinct imprint' a memory of a successful flight (not crashing). Please note: this imprint is NOT the teaching phase that momma bear will be doing (I guess that I am momma bear). This InsImp is simply an eNa 'chain' in the MT and with a similar airframe, the InsImp can be reused. Next we fly SARA again but we knock it out of the air with something like a frisbee. Using analytical computation we can '3dimensionalize' the frisbee and make it a little larger, like the size of an attacking bird or we can make it smaller, like the size of a bullet. We can create multiple InsImp models and fill our object database.
Now that we have some instinct in our MT eNa chain and a few models of objects, how do we 'teach' it something? In what way can we get our bear cub to explore a bit? Let's define exploring as an act of data collection until pain is received, when pain is removed exploration can continue. This is a very basic breakdown of a complicated system that happens throughout the natural world. Does a human baby have curiosity as an instinct? Putting things in it's mouth, staring intently at interesting objects or people making faces, interacting with it's environment. If it is not in pain (physical, hunger, emotional) then it is exploring, having fun. It is exploring because it CAN. So we can say that a child has a natural instinct to explore, in general and throughout human existence this is a fact.
The child does not know that the Golden Eagle is hunting it, but the momma bear knows. Does the bear cub need to be attacked to know that it is in danger? Learning or being taught about danger or pain is a large part of a child's growth. We teach children to protect themselves if we cannot. We teach children that the stove will burn their hand, we do not place the hand on the stove to demonstrate. In the case of the MT, 'parents' have an advantage. We have the opportunity to transfer data to SARA and basically 'teach' her what things are painful and that they should be avoided, if possible. Creating a concise 'Dangerous Situation Table' is required. Granted all terrible situations cannot be programmed, that is a variable. However, with enough DST data and fast enough processing, SARA should be able to look at the DST information and formulate a percentage based answer to the question "will this activity threaten me?".
With a threat model running on SARA's processor, data streaming to her in the form of DST information, model object availability and many sensors watching her surroundings, can we now place an InsImp in her chain that simply says to avoid death and collect sensor data? Combine that with a chain that allows her to boost or diminish her flight speed and also her distance to stationary or moving objects.
The real leap will be when she flies as fast and as close to danger as SHE sees fit or the way she will fly just as slowly as she can, because her cameras are pointed at that perfect burning sunset and she can feel it's rays charging her batteries through the solar array that is part of her airframe's skin.
http://phys.org/news/2016-04-humans-supercomputer.html
No comments:
Post a Comment