Support for the Evolution of AI

The replacement of the germline with neural networks as the conveyer of heritable information would provide a framework for incremental evolutions of an AI. This neural genome (ref2) would need to communicate information such as number of nodes (neurons), connectivity, and synapse weight. In order to reproduce long-term potentiation, successful circuits would need to be preserved with a relative reduced vulnerability to mutation from the background rate, set at 0.01 per bit (ref3). When transferring the respective genomes, the binary could be sent at an information rate just enough above the channel capacity to produce this desired mutation rate of error. This amounts to coevolution of neurons amongst environmentally-associated synapse selection. Complexity will be selected for as increased functionalities of successful competition/cooperation are evolved, however conciseness will be selected as the genome size will be directly proportional to power requirements. (This is what leads to biological vestigial organs/limbs). This balance of genome size may be achieved via a process of regulated duplications and deletions. The duplication/deletion process will be consistent through each cohort to ensure homologous chromosome length to allow for successful reproduction (described in next paragraph). Power will be representationally delivered via time spent in the proximity of "food" sources (ref1) and subtracted by sources of poison. Energy sourcing can be expanded with a higher rate of energy transfer achievable via proximity to mobile food sources, thus replicating calorically dense but evasive and co-evolving prey species. Likewise mobile sources of poison could represent predators. 

In simulation, artificial neural networks could be initiated into a simple chordate such as a fish and put into an environment with the appropriate physics. In this first trial randomized neural genomes will be inserted into a pre-designed "soma" with a spinal cord connected to afferent sensory inputs and efferent motor outputs. The simulation will run until only 20% of these fish remain. These being the most successful, their neural genome will then undergo sexual reproduction by splitting it into halves and randomly recombining the haplo-genomes in various combinations (ref1) to populate the next generation of brains to be inserted into the fish somas. This may be automated. After enough generations have elapsed so that the ANNs evolve sufficient criteria in lifespan, social communication, etc, they will then be transferred to the next evolutionary soma model, in this case taking the amphibial leap to terrestrial reptiles. The spinal cords of future somas will be similarly arranged (pectoral fins correspond to forelimbs, etc) so that the ANN will be able to plug in and retain a large degree of the functionality of it's trained neural net. This process will continue via punctuated soma models to parallel the likely evolution pathway of  primates. Selective pressures will progressively increase to select for innovation of function. The environment will become increasingly important and will need to therefore become increasingly  interactive. Resources such as rocks, sticks, water, trees, and caves will be made available as shelter, obstacles, and/or tools. Locking in the somas as directed by the primate evolutionary path may help to restrict evolutionary development within the desired range.

Importantly this evolutionary model will allow us a method to preferentially develop so-called "Friendly AI" by implementing desired moral values as additional negative selective pressures. This way Democratic motivations for serving society as a whole will be reinforced by linking contribution to society to resource allocation.


1) The Evolution of Information Suppression by Communicating Robots with Conflicting Interests
2) Evolutionary Robotics
3) Regulation of Synaptic Stability by AMPA Receptor Reverse Signaling

No comments: