| BoxesVisualDiscriminationExperiment |
Boxes Visual Discrimination Task, as described in: 'A Hypercube-Based Encoding for Evolving Large-Scale Neural Networks', Kenneth O. Stanley, David B. D'Ambrosio, and Jason Gauci (2009) http://eplex.cs.ucf.edu/publications/2009/stanley.alife09.html The task involves a visual field of 11x11 binary pixels (on/off, black/white). The field contains two boxes, one small and one large box that has an edge length 3x that of the smaller box. The goal is to identify the center pixel of the larger box. A single evaluation consists of 75 test cases with the two boxes randomly positioned. The root mean squared distance between selected and target pixels is scaled and translated to a range of 0-100, where 0 = no better than an agent randomly selecting pixels and 100 = perfect. In addition to this the range of output values is scaled to 0-10 and added to the final score, this encourages solutions with a wide output range between the highest activation (the selected pixel) and the lowest activation (this encourages prominent/clear selection). The problem domain view allows the performance of the best genome to be observed. The view also allows the HyperNEAT genome to be decoded using a range of visual field pixel resolutions beyond the 11x11 resolution that genomes are trained with. The ability to do this is a feature of HyperNEAT - the evolved genomes are Compositional Pattern Producing Networks (CPPNs). A CPPN defines the connection strength between nodes positioned in some euclidean space and therefore we can use a CPPN to produce neural networks with more nodes by increasing the number and density of nodes. |
