Neural network for modeling ecological and biological systems

• US 6,490,573 B1
• Filed: 04/11/2000
• Issued: 12/03/2002
• Est. Priority Date: 04/11/2000
• Status: Active Grant

• Current Assignee: Philip Chidi Njemanze
• Original Assignee: Philip Chidi Njemanze
• Inventors: Njemanze, Philip Chidi
• Primary Examiner(s): Black, Thomas
• Assistant Examiner(s): Holmes, Michael B.
• Application Number:
  US09/547,426
• Time in Patent Office: 966 Days
• Field of Search: 706/19, 706/12, 706/13, 706/22
• US Class Current: 706/19
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology

Abstract

A method of operating a neural network for ecological and biological system modeling having a plurality of hidden layer neurons said method comprising: a plurality of network inputs and at least one network output, said plurality of neurons, each receiving a plurality of inputs applied to the network, reproduces the network using a regression model, and compares the output values with given target values, and using the comparison and goodness of fit to set the learning rules. The network does not require repetitive training and yields a global minimum for each given set of input variables.

Claims

1. A method of operating a neural network for modeling ecological and biological systems having a plurality of hidden layer neurons, said method comprising the following steps:

(a) distributing network inputs to said hidden layer neurons as driving independent variables;

(b) said hidden layer neurons performing a user-specified regression model using the neuron weights as the dependent variable;

(c) said regression model at each step evaluates whether the fit of the model to the data has improved from the previous step, and calculating the loss function;

(d) said loss function is estimated using least squares estimation procedure aimed at minimizing the sum of squared deviations of the observed values for the independent variable from those predicted by the model stated as;

2. The method of claim 1 wherein the operation of the network is based on binary response in the form of a logit regression model;

wherein the predicted values for the dependent variable representing the weight value of the i th neuron will never be less than (or equal to 1 ), or greater than (or equal to

     1), regardless of the values of the independent variables;

wherein the equation is given by;

3. The method of claim 1 wherein the operation of said neural network is based upon responses that vary within a particular lower and upper limit using the general logistic regression model stated as:

4. The method of claim 2 wherein said binary response has the form of probit regression model stated as:

5. The method of claim 1 wherein the relationship between one or more independent input variables and a neuron weight as dependent variable changes over the range of the independent variables.

6. The method of claim 5 wherein the nature of the relationship applies a piecewise linear regression model stated as:

7. The method of claim 6 wherein the relationship applies the breakpoint regression model stated as:

8. The method of claim 1 wherein the nature of the relationship applies an exponential regression model stated as:

9. The method of claim 1 wherein the weights are directly or inversely proportional to the neuron output.

10. The method of claim 1 wherein the weights have established relationship with the neuron output.

11. The method of claim 1 wherein the network has a reproduction phase when all the weights must be already known.

12. The method of claim 1 wherein the network has a learning phase when one applies input values to the network, reproduces the network, and compares the output values with given target values.

13. The method of claim 12 wherein according to the result of this comparison, one modifies the weights using devised learning rules.

14. The method of claim 1 wherein the said neural network is implemented using integrated circuit.

15. The method of claim 1 wherein said neural network is implemented using a computer program.

16. The method of claim 1 with wide range applications in electronics, industry, commerce, medicine etc.

17. A method of operating a neural network for ecological and biological system modeling having a plurality of hidden layer neurons said method comprising the following steps:

(a) distributing network inputs as driving independent variables;

(b) said independent variables comprising ecosystem parameters selected on the basis of biological or physical relationships;

(c) said independent variables providing input to first layer of hidden neurons;

(d) said neurons comprising processes within the elements of the ecological and biological systems;

(e) said ecological and biological systems comprising bacteria, zooplankton, phytoplankton and hydrogeological features;

(f) said processes within the ecological and biological systems comprising neuron weights;

(g) said neuron weights having established biological relationship with neuron output;

(h) said output of the first layer neurons being fed as input to the second layer of hidden neurons;

(i) said second layer neurons generating input either to plurality of other hidden neuron layers or to the output neuron layer;

(j) said output neuron layer generating the total output of the network.

18. The method of claim 17 applying a user-specified regression model to a plurality of neurons in the hidden and output neuron layers.

19. A method of operating a neural network for ecological and biological system modeling having a plurality of hidden layer neurons said method comprising:

a plurality of network inputs and at least one network output, said plurality of neurons, each receiving a plurality of inputs applied to the network, reproduces the network using a current model, and compares the output values with given target values, said current regression model “

hierarchially relates”

such that the current model is identical to the previous model with the exception of an addition or deletion of one or more driving or independent variables to the previous model and using the comparison between the goodness of fit for the two models or difference to set the learning rules without need for repetitive training and yielding a global minimum for each given set of input variables.

20. The method of claim 19 wherein the state of the neural network is changing steadily in time;

wherein said network is simulated by a computer can alter its state only at discrete times t=0,1,2,3, . . . ;

wherein the time t=0 denotes the starting state of the network;

wherein given the state at the time t, the state at t+1 may be calculated.