Inference machine using adaptive polynomial networks
AssigneeGeneral Dynamics Corp., Pomona Div.
InventorBishnu P. Bhattacharya
Life of this patent
Estimating the term…
Abstract
An inference machine is described which comprises an adaptive polynomial network with the capability of implicitly inferring missing input information contained in time-dependent signals. The proposed network has wide applicability and is characterized by its ability to produce outputs within specified degrees of accuracy when, for example, only four out of ten input signals are available to the network. Specific applications include terminal guidance of a tactical missile when critical signals may be denied. A generic implementation is schematically illustrated.
Claims12
A method for constructing an inference machine for simulation of behaviour of a physical system within predetermined limits of accuacy when said inference machine has fewer independent inputs than said physical system, comprising the steps of:
- (a) designing an adaptive polynomial network having a layered structure of linearly transforming elements acting on time dependent signals arriving at a plurality of independent inputs, with successive layers of said structure making use of selected outputs of intermediate layers which have the most influence on said accuracy of simulation, and each said element having one or more numerical coefficients and a time delay of zero or greater associated therewith;
- (b) simulating said physical system with a digital computer;
- (c) amassing a "training" data set of simulated output values from the computer simulation of said physical system in response to input values;
- (d) running a training program of software instructions which operates on said training data set to produce a plurality of optimized values of said coefficients for said polynomials in said polynomial network, and to select a subset of values of said outputs of intermediate layers of said polynomial network which have the most important influence on said accuracy of simulation by said adaptive polynomial network of said physical system;
- and
- (e) implementing said adaptive polynomial network in hardware.
An inference machine comprising a polynomial network having an identical number of outputs as a system which said network models but a reduced number of independent inputs, wherein said network yields, within predetermined limits, the same output values as said system for the same input values of said reduced number of independent inputs, said network comprising;
- a noise-reducing layer of smoothing filters having a plurality of smoothed outputs and a plurality of inputs equal in number to said reduced number of independent inputs;
- a first layer of transformational elements connected in parallel, each said first-layer element having one of said smoothed outputs as an input and a plurality of outputs of sampled and stored values of said one smoothed input;
- a second layer of linear transformational elements connected in parallel, one for each said first-layer element, each said second-layer element comprising means of linearly transforming a predetermined number of said first-layer output values into an identical number of output values according to one specified linear transformational equation;
- and
- a plurality of sequential layers of transformational elements following said second layer, each said element in said sequential layers comprising means for linearly transforming a different predetermined combination of outputs from earlier layers into an output according to a particular specified linear transformational equation, with a last layer of said plurality of layers producing output values which are the same, within predetermined limits, as said system which said network models.
Dependencies are inferred from the claim text.