Neuroscience based machine intelligence models by Gary Gaulin, contact:

Monday, March 31, 2014

Grid Cell Attractor Network for place avoidance spatial navigation around Repelling border/boundary cell mapped hazards or barriers, Version 2

This update to the Grid Cell Network adds a Pulse command button and a more noticeable color coding to make it easier to study the AC component (of the field produced by all output connections of the attractor location staying active/on) where the alternating between sets of force vectors (violet lines showing force direction) average out to a more precise heading and encode a range of possible paths that can be taken, depending on behavior. Some animals prefer to follow walls/barriers while others prefer a more direct route like this model does. Code for repositioning the MyX,MyY location was much improved by using smaller steps calculated from local force vector field strengths, and converting back and forth between hexagonal grid network coordinates and Cartesian coordinates in its environment (required by the IDLab). The Attract and Repel arrays were eliminated by their 1 bit of data being stored in the uppermost 2 bits of the GridIn(X,Y) array byte, which also stores the 6 bits of grid field input from neighboring fields (N variable). Since the behavior of each field in response to these 8 bits of addressing input are all the same for each network X,Y location, N and Attract, Repel states the GridRAM(X, Y, N, A, R) array became simply GridRAM(GridIn(X,Y)) now addressed with only the GridIn(X,Y) byte. Training the GridRAM array for grid field behavior was then reducible to just four short lines of code, in the Initialize subroutine. The TimeStep code is now better optimized, and faster, even though these changes do not necessarily make it easier to understand how this Grid Cell Network model works, but might.

With compiled code for Windows:

No comments: