NeuroStrategy FAQ

14. Why does a strategy model not improve continuously during training?

Learning as such only makes up about 30% of the development process of a strategy model. The remaining 70% consist of analysis, testing and adjustment.

The trainer in the system constantly supervises the mesh, the "electronic brain", which is being developed. He is able to ascertain, for example, whether the answer to a particular question was found through abstraction and combinatorial analysis, or whether the mesh had simply stored the answer. The latter would contradict the goal of the developmental process, the formation of intelligence. Accordingly, the trainer can weaken or even erase those areas of the mesh, which fulfill a pure memory function and do not actively contribute to an intelligent solution. This makes learning more difficult for the system. The system forgets those facts, which earlier it had simply memorized by heart. This forces the mesh to use its resources intelligently. In the meantime this can lead to a reduction in training performance, since the mesh must first find and develop new approaches towards a solution.

A further reason why there are setbacks during training lies in the manner in which a mesh develops. It can only feel its way towards a particular solution through a gradual change of the strength of the signal transmission of a synapse in the electronic brain. Thus, a synapse that currently acts in a signal-amplifying way cannot in the next training run exercise an opposite function. Instead, the signal transmission first has to be gradually reduced to 0. The approach to a solution thus resembles a journey full of climbs and descents, where the peak that is the destination is only reached via a walk through the valley in between. During the crossing of such a valley, performance can likewise decrease.