Shortcomings with using edge encodings to represent graph structures

摘要

There are various representations for encoding graph structures, such as artificial neural networks (ANNs) and circuits, each with its own strengths and weaknesses. Here we analyze edge encodings and show that they produce graphs with a node creation order connectivity bias (NCOCB). Additionally, depending on how input/output (I/O) nodes are handled, it can be difficult to generate ANNs with the correct number of I/O nodes. We compare two edge encoding languages, one which explicitly creates I/O nodes and one which connects to pre-existing I/O nodes with parameterized connection operators. Results from experiments show that these parameterized operators greatly improve the probability of creating and maintaining networks with the correct number of I/O nodes, remove the connectivity bias with I/O nodes and produce better ANNs. These results suggest that evolution with a representation which does not have the NCOCB will produce better performing ANNs. Finally we close with a discussion on which directions hold the most promise for future work in developing better representations for graph structures.