The research experience I've been accepted to for next semester is a part of ongoing research into nuturing in machine learning as well as strongly related topics.
Looking into some of their prior work, they've done some very interesting experiments. For example, they evolved complex behavior through artificial neural networks. The goal of the robots is to turn on a light and then stay under the light for as long as possible, simulating gaining energy, thus survivability. The robots get eighteen inputs, determined by seeing one of three possible colors (for the switch, the light, and the light off) in different regions of it's simple camera, and has two outputs that control left and right motors. The input is mapped to output through a simple neural network, no hidden layers. The robots start from absolutely no knowledge, moving about randomly. The robots that are most 'fit' (most time under the light in their run) at the end of the trial get to reproduce, with their genes (mappings/weights) slightly mutated. Over enough generations, the robots being very efficient at turning on the light and going straight to it.
Here's where it gets a little more interesting: The investigation into parent-child nurturing. As expected, the parent accomplishes its duty by turning on the light, and the child does its part by heading immediately to the light while it's still off as to be ready. However, it's not uncommon for the child to fall off the arena. So an unexpected behavior emerges in some 'families': After turning on the light, the parent actually maneuvers around to the other side of the arena to BLOCK the child from falling off. Pretty incredible behavior for such a simple system.
A key motivation for this research is illustrated in Figure 1 of the first attached document.
Which leads me to this video of a true neural network at work:
[ame="http://www.youtube.com/watch?v=1QPiF4-iu6g"]Rat Brain Robot - YouTube[/ame]
Some added explanation: http://neurobonkers.com/tag/rat-brain-robot/
Looking into some of their prior work, they've done some very interesting experiments. For example, they evolved complex behavior through artificial neural networks. The goal of the robots is to turn on a light and then stay under the light for as long as possible, simulating gaining energy, thus survivability. The robots get eighteen inputs, determined by seeing one of three possible colors (for the switch, the light, and the light off) in different regions of it's simple camera, and has two outputs that control left and right motors. The input is mapped to output through a simple neural network, no hidden layers. The robots start from absolutely no knowledge, moving about randomly. The robots that are most 'fit' (most time under the light in their run) at the end of the trial get to reproduce, with their genes (mappings/weights) slightly mutated. Over enough generations, the robots being very efficient at turning on the light and going straight to it.
Here's where it gets a little more interesting: The investigation into parent-child nurturing. As expected, the parent accomplishes its duty by turning on the light, and the child does its part by heading immediately to the light while it's still off as to be ready. However, it's not uncommon for the child to fall off the arena. So an unexpected behavior emerges in some 'families': After turning on the light, the parent actually maneuvers around to the other side of the arena to BLOCK the child from falling off. Pretty incredible behavior for such a simple system.
A key motivation for this research is illustrated in Figure 1 of the first attached document.
Which leads me to this video of a true neural network at work:
[ame="http://www.youtube.com/watch?v=1QPiF4-iu6g"]Rat Brain Robot - YouTube[/ame]
Some added explanation: http://neurobonkers.com/tag/rat-brain-robot/
Attachments
Last edited: