Awareness without Neural Networks: Achieving Self-Aware AI via Evolutionary and Adversarial Processes

EasyChair Preprint 4317

Abstract

A key difficulty in achieving self-aware artificial intelligence (AI) is the achievement of epistemological knowledge, i.e. a machine that “knows what it knows” and “knows what it does not know” with respect to some model of itself or its surroundings. Given a nonlinear dynamical system with known algebraic structure expressible as differential equations, with sensors able to create a time-series of measurements of sufficient variables to create a suitable partial state vector, then novel forms of evolutionary machine learning and adversarial processes are sufficient to create a form of AI that is “aware” of its knowledge set regarding this system, and can use a form of differential Game Theory and adversarial processes to “think” about its knowledge set to address ambiguities and achieve objectives, including moving beyond its original training data. This may itself constitute a form of “self-awareness”.

Results from successful use of these techniques in medical and engineering problems are outlined. This AI architecture does not involve neural networks or their derivative architectures, but instead is inspired by evolutionary ecosystems. Implications for self-aware operating systems are discussed.

Keyphrases: Adversarial process, Differential Game Theory, Epistemological, Evolutionary, Evolutionary Machine Learning, Genetic Algorithm, control strategy, dynamics, ecosystem, game, learning, machine learning, nonlinear dynamical system, textured evolutionary algorithm

@booklet{EasyChair:4317,
  author    = {Nigel Greenwood and Brruntha Sundaram and Alex Muirhead and James Copperthwaite},
  title     = {Awareness without Neural Networks: Achieving Self-Aware AI via Evolutionary and Adversarial Processes},
  howpublished = {EasyChair Preprint 4317},
  year      = {EasyChair, 2020}}