Network that learns what it actually does

• US 5,572,627 A
• Filed: 03/23/1995
• Issued: 11/05/1996
• Est. Priority Date: 11/22/1993
• Status: Expired due to Fees

• Current Assignee: Robert A Brown
• Original Assignee: Robert A Brown
• Inventors: Brown, Robert A.
• Primary Examiner(s): Davis, George B.
• Application Number:
  US08/408,992
• Time in Patent Office: 593 Days
• Field of Search: 395/11, 395/24, 395/425, 395/20
• US Class Current: 706/14
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/063 using electronic means

Abstract

A self-learning network with: (a) one or more sensor(s) in one or more a sensor unit(s) with a plurality of memory cells that forward-select values of one or more intermediate variable(s) I₁, J₁, . . . according to their stored values and the conditions identified by its sensor(s) in a given sensing and action period, (b) with or without one or more intermediate units with a plurality of bidirectional memory cells that forward-select a value of one or more intermediate output variable(s) I₂, J₂, . . . according to their stored values and the value(s) of the intermediate input variable(s) I₁, J₁, . . . , and (c) one or more actuator(s) in one or more actuator unit(s) connected to one or more intermediate variable(s) I₁, J₁, . . . or I₂, J₂, . . . with a plurality of bidirectional memory cells that forward-select values of one or more actuators according to their stored values and the value(s) of its intermediate variable(s) I₁, J₁, . . . or I₂, J₂, . . . The bidirectional memory cells then back-select a value of each intermediate co-variable I₁ '"'"', J₁ '"'"', or I₂ '"'"', J₂ '"'"', . . . according to their stored values and the actions that actually occur within each specific sensing and action periods. The stored values are reduced in the forward-selecting memory cells and, and the stored values are increased in the back-selecting bidirectional memory cells, and in the memory cells at the intersection of the back-selected value(s) of the intermediate co-variable I₁ '"'"', J₁ '"'"', . . . and the conditions identified by its sensor(s) in a given sensing and action period.

Claims

1. A bidirectional memory means useful in a self-learning machine with a network of memory means, said bidirectional memory means operating in a sequence of sensing and action periods, said bidirectional memory means comprising:

a. a timing means to herald a beginning and an end of a sensing and action period of predetermined duration, said sensing and action period being within said sequence of sensing and action periods, and

b. at least one of an input interface means to identify an extant value of said input interface means near a beginning of said sensing and action period, and

c. at least one of an output interface means to make a production of a forward-selected value of said output interface means within said sensing and action period, and

d. said output interface means also to identify an extant value of said output interface means after a predetermined delay from said production of said forward-selected value of said output interface means, and

e. said input interface means also to make a production of a back-selected value of said input interface means within said sensing and action period, and

f. said bidirectional memory means containing at least one of a bidirectional submatrix means made up of bidirectional memory cell means, wherein each said bidirectional submatrix means is connected to one said input interface means and one said output interface means uniquely, wherein

g. said bidirectional memory means is disposed to record a historical probability between said extant value of said input interface means and said extant value of said output interface means during said sequence of sensing and action periods, and

h. said bidirectional memory means also is disposed to make a forward-selection of said forward-selected value of said output interface means in said sensing and action period, said forward-selection being made on the basis of said a highest of said historical probability, and

i. said bidirectional memory means also is disposed to make a back-selection of said back-selected value of said input interface means in said sensing and action period, said back-selection being determined by said highest of said historical probability, whereby

j. said bidirectional memory means is a nodal unit means in a self-learning machine with a nodal network, and said bidirectional memory means is used in an actuator unit means in a self-learning machine with a duplex network, and a self-learning machine with a nodal network.

2. The bidirectional memory means of claim 1 with at least one of a bidirectional submatrix means made up of a plurality of a bidirectional memory cell means, wherein each said bidirectional submatrix means is connected to one of an input interface means and one of an output interface means uniquely, more specifically wherein:

a. one of a forward-selecting bidirectional memory cell means in each said bidirectional submatrix means contributes to a forward-selection of a forward-selected value of said output interface means, where said forward-selecting bidirectional memory cell means belong to a forward-selecting set, said forward-selecting set being connected to said forward-selected value of said output interface means, and where said forward-selecting set also belong to a forward-selecting energized set of said bidirectional memory cell means, said forward-selecting energized set being connected to an extant value of said input interface means, and wherein

b. a lowest sensitivity forward-selecting bidirectional memory cell means in each said forward-selecting set is higher than a lowest sensitivity of a non-selecting bidirectional memory cell means in any of a non-selecting set, each said non-selecting set being connected to an unselected value of said output interface means, said non-selecting bidirectional memory cell means also being in said forward-selecting energized set, and where

c. a sensitivity of each said forward-selecting bidirectional memory cell means is subjected to a reduction, said reduction being equal to a sensitivity of each said forward-selecting bidirectional memory cell means multiplied by a predetermined constant, said predetermined constant being in a range greater than zero and less than one, and

d. one of a back-selecting bidirectional memory cell means in each said bidirectional submatrix means contributes to a back-selection of said back-selected value of said input interface means, where said back-selecting bidirectional memory cell means belong to a back-selecting set of bidirectional memory cells means connected to said back-selected value of said input interface means, where said back-selecting bidirectional memory cell means also belong to a back-selecting energized set of bidirectional memory cell means connected to an extant value of said output interface means, and wherein

e. a lowest sensitivity back-selecting memory cell means in each said back-selecting set is higher than a lowest sensitivity non-selecting of a non-selecting bidirectional memory cell means in any of a non-selecting set, where each said non-selecting set is connected to an unselected value of said input interface means, said non-selecting bidirectional memory cell means also being in said back-selecting energized set, and where

f. a sensitivity of each said back-selecting bidirectional memory cell means in each said bidirectional submatrix is subject to an increase, said increase being equal to a difference between a maximum possible sensitivity of said bidirectional memory cell means and said sensitivity of said bidirectional memory cell means, said difference being multiplied by said predetermined constant, wherein

g. said bidirectional memory means forward-select a forward-selected value of each said output interface means, said forward-selected value having occurred most consistently with said extant value of each of said input interface means in a sensing and action period, whereby

h. said bidirectional memory means is a nodal unit means in a self-learning machine with a nodal network, and said bidirectional memory means is used as a bidirectional memory means in an actuator unit means in a self-learning machine with a duplex network and a self-learning machine with a nodal network.

3. The bidirectional memory means of claim 2 with at least one of an input interface means and at least one of an output interface means, with an addition of at least one of an actuator means, one said actuator means being connected to one said output interface means uniquely, said bidirectional memory means and said actuator means forming an actuator unit means wherein:

a. said bidirectional memory means makes a forward-selection of a forward-selected value of said output interface means, said forward-selection being made near a beginning of a sensing and action period, said forward-selection being made according to an extant value of each of an input interface means and a sensitivity stored in said bidirectional memory means, and wherein

b. said forward-selected value of each of said output interface means causes each said actuator means to make an attempt to assume a forward-selected value of said actuator, said attempt being made after a predetermined action delay from said beginning of said sensing and action period, said attempt being made for a predetermined action period, and where

c. said actuator means makes an attempt to produce said forward-selected value of said actuator means during an action period within said sensing and action period, and

d. said output interface means measures an actual value of each said actuator means during a feedback period within said sensing and action period, said feedback period occurring after said action period, and

e. said actuator unit means also has an actuator brake means to make a restraint upon said actuator means, said restraint being made except when said actuator means makes said attempt to produce said forward-selected value during said action period, and

f. said bidirectional memory means makes a back-selection of a back-selected value of said input interface means, said back-selection being made during a feedback period within said sensing and action period, said back-selection being made according to said actual value of each of said actuator means and a sensitivity stored in said bidirectional memory means, where

g. said forward-selected value has occurred most consistently with said extant value of said input interface means in a previous plurality of sensing and action periods, and said back-selected value of said input interface means is most likely to forward-select said actual value of said actuator means in said sensing and action period, whereby

h. said actuator unit means is useful in a self-learning machine with a network of memory means.

4. The actuator unit means of claim 3 with an addition of a sensor unit means, said sensor unit means being connected to said actuator unit means by an intermediate interface means, said addition forming a self-learning machine with a duplex network, said duplex network comprising:

a. said sensor unit means with least one of a submatrix means, said submatrix means made up of a plurality of a memory cell means, wherein each said submatrix means is connected to one of a sensor means and one of an intermediate interface means uniquely, and

b. one of a forward-selecting memory cell means in each of said submatrix means contributes to a forward-selection of a forward-selected value of said intermediate interface means, where said forward-selecting memory cell means belong to a forward-selecting set of said memory cells means, said forward-selecting set being connected to said forward-selected value of said intermediate interface means, and where members of said forward-selecting set also belong to a forward-selecting energized set of said memory cell means, said forward-selecting energized set being connected to an extant value of said sensor means at a beginning of a sensing and action period, and wherein

c. a lowest sensitivity forward-selecting memory cell means in each said forward-selecting set is higher than a lowest sensitivity non-selecting memory cell means in any of a non-selecting set of said non-selecting memory cell means, each said non-selecting set being connected to an unselected value of said output interface means, said non-selecting memory cell means also being in said forward-selecting energized set, and where

d. a sensitivity of each said forward-selecting memory cell means is subjected to a reduction, said reduction being equal to a sensitivity of each said forward-selecting memory cell means multiplied by a predetermined constant, said predetermined constant being in a range greater than zero and less than one, and

e. said actuator unit means makes a forward-selection of a forward-selected value of each of at least one of an actuator means according to each said forward-selected value of each of said intermediate variable means, and

f. said actuator unit means makes a back-selection of a back-selected value of said intermediate interface means, said back-selection being made according to a measured value of said actuator means, and

g. a sensitivity of each of a feedback-selected memory cell means in said sensor submatrix is subjected to an increase, said feedback-selected memory cell means being connected to said extant value of said sensor means and said back-selected value of said intermediate interface means, said increase being equal to a difference between a maximum possible sensitivity of said memory cell means and said sensitivity of said memory cell means, said difference being multiplied by said predetermined constant, wherein

h. said self-learning machine with a duplex network selects a value of each said actuator means that have most consistently occurred with said extant value of each said sensor means in a sensing and action period, whereby

i. said duplex network of said self-learning machine requires fewer memory cell means when fewer sensor/actuator relations are required by a control task, said self-learning machine with a duplex network also being self-organizing.

5. The self-learning machine with a duplex network of claim 4, further including a digitizing means, forming a digitized self-learning duplex machine, where in a sensing and action period:

a. an extant value of a sensor means is decomposed by an encoding means into a unique combination of values of a plurality of an aggregate input variable means of a sensor unit means, and

b. said sensor unit means forward-selects a value of each of an intermediate interface means according to said unique combination of values of said plurality of said aggregate input variable means, said value of each of said intermediate interface means having been back-selected most consistently with said unique combination of values of said plurality of said aggregate input variable means within said sensing and action period, and

c. said actuator unit means forward-select a unique combination of values of a plurality of an aggregate output variable means according to said forward-selected value of each of said intermediate interface means, said unique combination of values of said plurality of said aggregate output variable means having occurred most consistently with said forward-selected value of each of said intermediate interface means, and

d. a selected value of an actuator means is synthesized by a decoding means from said unique combination of values of said plurality of said aggregate output variable means of said actuator unit means, and

e. said actuator means attempts to produce said selected value of said actuator means, and

f. a measured value of said actuator means is decomposed by an encoding means into a unique combination of values of a plurality of aggregate output co-variable means of said actuator unit means, and

g. said actuator unit back-selects said back-selected value of said intermediate interface means according to said unique combination of values of said plurality of said aggregate output co-variable means, said back-selected value of said intermediate interface means having occurred most consistently with said unique combination of values of said plurality of said aggregate output co-variable means, where

h. said digitized self-learning duplex machine attempts to produce said selected value of said actuator means, said selected value of said actuator means having occurred with said extant value of said sensor means most consistently in each said sensing and action period, whereby

i. said digitized self-learning duplex machine requires fewer memory cell means than an undigitized self-learning machine, said digitized self-learning machine also being self-organizing.

6. The self-learning machine with a duplex network of claim 4 with an addition of a nodal unit means, said addition forming a self-learning machine with a nodal network, said nodal unit means operating in a sequence of sensing and action periods, said nodal unit means comprising:

a. a timing means to herald a beginning and an end of a sensing and action period of predetermined duration, said sensing and action period being within said sequence of sensing and action periods, and

b. at least one of an input interface means to identify an extant value of said input interface means near a beginning of said sensing and action period, and

c. at least one of an output interface means to make a production of a forward-selected value of said output interface means within said sensing and action period, wherein

d. said output interface means also to identify an extant value of said output interface means after a predetermined delay from said production of said forward-selected value of said output interface means, and

e. said input interface means also to make a production of a back-selected value of said input interface means within said sensing and action period, and

f. said nodal unit contains at least one of a bidirectional submatrix means made up of bidirectional memory cell means, wherein each said bidirectional submatrix means is connected to one said input interface means and one said output interface means uniquely, wherein

g. said bidirectional submatrix means is disposed to record a historical probability between said extant value of said input interface means and said extant value of said output interface means during said sequence of sensing and action periods, and

h. said bidirectional submatrix means also is disposed to make a forward-selection of said forward-selected value of said output interface means in said sensing and action period, said selection being made on the basis of a highest of said historical probability, and

i. said bidirectional submatrix means also is disposed to make a back-selection of said back-selected value of said input interface means in said sensing and action period, said back-selection being determined by said highest of said historical probability, and

j. said input interface means of said nodal unit means is connected to an intermediate interface means of at least one of a sensor unit means, and said output interface means is connected to an intermediate interface means of at least one an actuator unit means, where

k. said actuator unit means makes an action attempt in response to an extant value of said sensor means in said sensing and action period, said action attempt having been carried out with said highest of said historical probability with said extant value of said sensor means in each previous said sensing and action period, whereby

l. said self-learning machine with a nodal network can connect multiple sensor unit means to multiple actuator unit means with fewer interconnections, fewer of a memory cell means, and fewer said bidirectional memory cell means, said self-learning machine also being self-organizing.

7. The self-learning machine with a nodal network of claim 6, additionally including digitized connections between each of a sensor unit means and a sensor means, and digitized connections between each of an actuator unit means and an actuator means, providing a digitized self-learning machine with a nodal network, said digitized self-learning machine with a nodal network also being self-organizing.