Ai Dreams Forum
AI Dreams => General Chat => Topic started by: ranch vermin on March 14, 2018, 08:07:16 pm
-
So evolution based artificial intelligence uses noise to apply flukes/mutations to novel situations. nothing is done with purpose.
And this is state of the art since Karl Sims!!!! Theres got to be something new to generate behaviour to novel patterns that isnt just this.
If anyone figures it out, id keep it to yourself and make it a long drawn out complex hard to read implementation, because its a nifty secret that might power a singularity.
-
I've thought about it myself before. It would get exponentially closer. It may also get caught in a local valley.
-
For me i take noise form some where and then add it to input data to a RNN to get slightly different output data.
AGI mind branching system selects most wanted item fist. Noise mixed in will give the other lessor wanted items. This is
forcing the RNN to look 5 items kinda things at the same time. IF it causes a improvement the the noise plus the input nn data are
added together and saved.
-
Yes, you guys have an idea about this. Ive thought it through a little more, but I need to get my system up and running before ill see the truth of this, because it depends on what your applying the noise to.
So noise is what we have been using as a collective group of scientists to add novelty to the machine to keep it exploring new areas. I think systems like this cant jump to a conclusion, they can only hone what they already know. so to get a system to work out how to use tools and speak (which i think go together) its in the symantics and how your gathering knowledge into your system. and how you are using the knowledge.
My best idea after working this lot out, is I think you should have a general symbolic playback which gives the droid ideas, and a separate more detail playback for attempting them for real.
And that way itll jump to more distant moments, and its search is more free to try new things out.
-
Isn't this the best way? Scroll down and see the cool GIF.
https://en.wikipedia.org/wiki/Simulated_annealing
-
The devils in the details how it isnt true intelligence, and is just primitive.
Youll have to get up and running before youll know why it actually isnt intelligence yet. It depends on how you do it too, what results youll get.
-
Another i am using noise is on the weight of a very small NN. Fist weigh are 100 percent noise. This Little NN is detector with one output.
Not a binary output. It sample data by moving around. If it gets 22 percent of max activation it stop on that data spot. Clones that NN and add a
little bit of noise to the original weights and saved. This a try to improve the detection. This is a high speed mutational genetic unsupervised
algorithm.
The clone or child NN is pared off into adjacent area . So it will build a SOM, self organizing map, at the same time.
AI That Creates AI. Aaround 4.44 time into vid:
https://www.youtube.com/watch?v=hVv68aHYSs4&t=475s
-
VINE: An Open Source Interactive Data Visualization Tool for Neuroevolution:
https://eng.uber.com/vine/
-
Computing With Random Pulses Promises to Simplify Circuitry and Save Power:
https://www.spectrum.ieee.org/computing/hardware/computing-with-random-pulses-promises-to-simplify-circuitry-and-save-power
-
Simulated annealing is widely used to place and route FPGA designs, but Vivado (possibly the most recent/advanced tool in that field) has moved away from that, and instead uses a deterministic, analytical approach that just computes a cost function and finds the global minimum. Apparently simulated annealing doesn't scale well ... if you're dealing with a large number of elements that need to be placed, and/or a very congested design (which means more elements competing for the available placement locations) runtime and/or quality of results can suffer badly.
I dunno how relevant FPGA place and route is to the types of problems an AGI needs to solve, but I thought I'd throw that out there.