Artificial Intelligence #
Reading Chapter 7 (Artificial Creativity) of “The Beginning of Infinity” by David Deutsch gave me some ideas for things to try out WRT creating some kind of artificial intelligence.
Modelling the Intelligence #
The AI could be composed of “neurons” in a graph data structure. Each neuron would have multiple other neurons connected to it’s “input” side, and connect to one other neuron as it’s “output”. It would carry a single “fire/no fire” signal during every tick of time. The signal would be “fire” if during the last tick more than half (or some percent) of the input signals were “fire”, otherwise it would be “no fire”.
Neurons would be grouped arbitrarily into sub-networks with relatively few inputs/outputs compared to the interconnections within them.
Evolution #
The model would evolve using a random application of the following changes:
- A random neuron would move its output side to another random neuron, with a preference to move its output to neurons “close” to it, as judged by distance in the graph.
- A random neuron would be replaced by a copy of another neuron group with the same number of inputs and a single random output. This is simulating “creativity”, or the combining of disparate ideas. Another way to do this would be to connect a single neuron’s output to the input of another group, although this may lead to situations where that new group becomes ineffective due to overloading of random inputs.
Evolution can happen to a network during its normal running (it can learn on the fly).
It might be interesting to think more about the difference between “learning” and “evolution” - perhaps at a certain point only “learning” will be allowed, and not “evolution”. To start out with it probably makes sense to not demarcate these.
Input/Output #
The entire network would have a very simple output (perhaps even just one binary signal), and few/no inputs.
Starting Points #
I think the nervous system of C. elegans has been completely mapped, so maybe copying that would be a good place to start?
Objectives #
I don’t really know what to do here (how to bias the network to evolve in a specific direction, or how to reward it for making “good” changes).
Prior Art #
Categories: Thought Experiments