GB1317053A - Search-minimized optimum processor - Google Patents

Abstract

1317053 Trainable processors TEXAS INSTRUMENTS Inc 11 June 1970 [30 June 1969] 28363/70 Heading G4R In a trainable processor, similar to that described in Specification 1,264,387, training functions developed in the course of training are stored in a random access memory. For each set of input signals representing a training function key signals are generated to designate node locations defining a branch of a tree storage array. The key signals and training function are stored in the tree array in the random access memory. This arrangement reduces the number of storage locations required since a storage location is assigned as needed for each training function, whereas in the prior processor a location is provided uniquely for each of the possible combinations of all the quantization levels of the signals used in training. During training the processor is fed with a succession of sets of input signals u i and a corresponding series of signals representing the desired response z i of the processor to the signals u i . The signals are quantized and used to develop key signals which address locations in the memory. Quantized signals representing the output x i-1 of the processor in response to prior training signals are also used in the development of the key signals. The selected locations' contents are adjusted in accord with z i and u i . The branch in the tree storage is defined by the key signal, and this signal is compared with keys representing previously defined branches, if any. If such a path has previously been defined the content of the last storage location in the path is adjusted to correspond to the present input signal set. The adjustment to the content takes the form of adding the value of the desired response z i to the current content to provide a result which replaces the current content. The number of times that the location has been adjusted is accumulated, and the processor output signal x i is determined by dividing the adjusted content by this number. If no corresponding path has been previously defined the key signal defines a new path in the tree. During execution, the processor provides an output signal x i in response to an input u i , x i being obtained from the tree storage. If the processor encounters an untrained input it replaces this input by the value of the previous output.