Relation learning with bar charts

The paper reports on the work in progress on developing a machine learning method that is inspired on how a human operator might visually try to find relations between a large number of incoming data streams originating from sensors and controllable actuators. A large number of plot-like data structures of various dimensions are periodically generated for the combinations of given data streams and searched for regularities. Controllable variables are manipulated in some random or systematic way until the observed regularities allow for more intelligent behavior (if required). Information about interesting (as defined by the task at hand) relationships is extracted from the data structures, and human-readable control rules for the system are formed. For most nontrivial applications the method can be computationally very expensive, but the potential for parallelizability and for smart optimizations together with the continual increase of capabilities of the off-the-shelf computers suggest that the method will be feasible already in the nearest future.