Value Cell Encoding Strategies
John P. Sullins · Defense Technical Information Center (DTIC) · 1985
In many application areas, particularly in the biological sciences, there is the need to store several values of variables. Given a finite precison, one can store these values in Nk explicit cells, referred to as value cells, in a k-dimensional space of grain N. Typically, the number of values that must be stored is a very small fraction of the total number specified by the grain of the multidimensional space. This leads to data structuring that reduces the number of explicit cells required for a given level of accuracy. One idea is coarse coding, intersection of larger, coarser grained cells. Coarse coding has been shown to reduce the number of cells required by a factor of 1/Dk-1 where D is the diameter of the coarse cell in units of fine grained cells. This intuitively appealing idea in fact involves many subtle tradeoffs that are the focus of this paper. Coarse coding is shown to be independant of the isotrophy of the cells and superior to simply reducing the grain of the representation space. Loss of information due to the possibility of some fine grained cells sharing some of the same coarse grained cells and due to uncertainty in the input and translation of data is examined.