release 0.0.4
NarConSimpleWorld
A NAR(in this case ONA) can control a agent in a virtual world (simplified pong or a simplified shoot'em'up game) which is rendered as a virtual map.
The vision pipeline does:
- convolution for edge detection (8 directions)
- sampling of edges as symbolic points (by threshold) into set X
- recognition of points as lines (not strictly necessary)
- rendering of points of lines back to black/white image
- cropping of rendered image for unsupervised classifier C by proposal positions (derived from X)
generation of region proposals:
- cluster points from set X by proximity
- clusters are proposals, used for classifier C, have center and extend. Proposals are represented as AABB's
classifier C:
- unsupervised (as usual)
- online learning(as usual)
- try to find closest match by distance of stored classes
- if not found -> create new class with new class id
- if found -> return class id