[2017.10.31] From Genetic algorithm to Neuro-Evolution algorithm

[unizard]

Presneter: Teawon Kim Slide: 20171023_neuroevolution.pdf

Title: From Genetic algorithm to Neuro-Evolution algorithm, as an alternative to the gradient-based learning paradigm

Abstract: I will briefly review the genetic algorithm and its extension to the neuro-evolution algorithm. In the neuro-evolution literature, there is no need to compute gradients and a differentiable objective so that it can be a possible alternative to a model whose gradients cannot be computed explicitly (e.g. polity gradient in the reinforcement learning). I will present several neuro-evolution algorithms currently published, and we will discuss whether these algorithms can be survival in the future or not.

"The evolution of structural organization may be key to evolving true artificial general intelligence", Joost Huizinga, Evolving AI Lab., University of Wyoming

ref: K. O. Stanly et. al., Evolving Neural Networks through Augmenting Topologies, MIT Press, 2002 T. Salimans et. al., Evolution Strategies as a Scalable Alternative to Reinforcement Learning, arXiv preprint, 2017 E. Real et. al., Large-Scale Evolution of Image Classifiers, ICML, 2017 R. Miikkulainen et. al., Evolving Deep Neural Networks, arXiv preprint, 2017 https://blog.openai.com/evolution-strategies/ OpenAI research blog

[unizard]

Idea

Human inspired Feature...의 성장... 사람이 특징점이라고 정의했던 영역들이 과연 interesting points 일까? 에 대한 부분들. 이런 진화적 관점에서의 재 설계는 어떠할까