Spectrum-diverse Unified Neuroevolution Architecture

Spectrum-diverse Unified Neuroevolution Architecture (SUNA) is an evolutionary algorithm developed by Danilo Vasconcellos Vargas and Junichi Murata for evolving neural networks in which both the topology as well as parameters of the network are developed. SUNA aims to work completely autonomously, learning any kind of problem without prior knowledge or the need of an expert. To achieve this, two problems needed to be tackled:

1. A powerful representation that is flexible enough to model any kind of problem need to be developed.
2. Consequently, to develop a powerful and complex representation a specialized learning algorithm is needed.

SUNA tackles these two problems with the following key features respectively:

1. Unified Neural Model – To develop a powerful representation that is capable of modelling a wide range of problem classes, a novel neural model called Unified Neural Model is proposed that unifies most neural network features from the literature into one representation.
2. Spectrum-Diverse Neuroevolution – To develop such complex representations, neuroevolution is used with a new concept that calculates the spectrum of candidate solutions based on their characteristics and use it to keep the diversity. This enable high dimensional structures to be compared efficiently and scale well with their size.

Results

SUNA either surpassed or equalled NEAT on all 5 different problem classes.^[1] Ablation tests showed that this was possible because SUNA can adapt the representation depending on the problem. For example, employing neuromodulation when the problem needs it.

References

1. Danilo Vasconcellos Vargas, Junichi Murata, IEEE Transactions on Neural Networks and Learning Systems

External links

• Physis-Shard on GitHub