Discrete-time optimal control by neural network

Claim 2

A neural network for determining according to a quadratic cost function optimal control inputs for a linear discrete-time process at M sampling times, the process being characterized by and r control variables, comprising: N═(p+r) M neurons, a distinct neuron being assigned to represent the value of each state variable at each sampling time and a distinct neuron being assigned to represent the value of each control variable at each sampling time; an input for each neuron; an output for each neuron; an external input connected to the neurons representing the initial values of each state variable and each control variable, the external input having a value determined by a modified quadratic cost function J' defined as: ##EQU8## where A is a p×p state transition matrix, B is a p×r control distribution matrix, X k is a p×1 state vector, u k is an r×1 control vector at the sampling time instant k, Q and R are symmetric weighting matrices, M is the number of time steps, λ is a finite constraint strength, and I is the identity matrix; and a connection between the output of each neuron and the input of each other neuron in the network, the strength of each such connection being determined by the modified cost function, such that running the neural network for a sufficient time to minimize the cost function will produce optimum values for each control variable at each sampling time.

Claim 3

The network of claim 1, wherein each neuron further comprises an analog amplifier.

Claim 4

The network of claim 2, wherein the network further comprises a Hopfield network, where the change with time of the state variables s i in the network is governed by the coupled nonlinear differential equations: ##EQU9## where C i is the ith neuron's input capacitance; R i is the ith neuron's resistance, T ij is the connection strength between the ith and jth neurons, f(s) is a smooth, monotonically nondecreasing function, and I i is a bias current.

Claim 5

A Hopfield neural network for determining according to a quadratic cost function optimal control inputs for a linear discrete-time process at M sampling times, the process being characterized by p state variables and r control variables, where the change with time of the state variables s i in the network is governed by the coupled nonlinear differential equations: ##EQU10## where C i is the ith neuron's input capacitance; R i is the ith neuron's resistance, T ij is the connection strength between the ith and jth neurons, f(s) is a smooth, monotonically nondecreasing function, and I i is a bias current, comprising: N═(p+r) M neurons, each neuron comprising an analog amplifier, a distinct neuron being assigned to represent the value of each state variable at each sampling time and a distinct neuron being assigned to represent the value of each control variable at each sampling time; an input for each neuron; an output for each neuron; an external input connected to the neurons representing the initial values of each state variable and each control variable, the external input having a value determined by a modified quadratic cost function J': ##EQU11## where R is a p×p state transition matrix, B is a p×r control distribution matrix, x k is a p×1 state vector, u k is an r×1 control vector at the sampling time instant k, Q and R are symmetric weighting matrices, M is the number of time steps, λ is a finite constraint strength, and I is the identity matrix; and a connection between the output of each neuron and the input of each other neuron in the network, the strength of each such connection being determined by the modified cost function, such that running the neural network for a sufficient time to minimize the cost function will produce optimum values for each control variable at each sampling time.

Claim 6

A Hopfield neural network for determining according to a quadratic cost function optimal control inputs for a linear discrete-time process at M sampling times, the process being characterized by p state variables and r control variables, where the change with time of the state variables s i in the network is governed by the coupled nonlinear differential equations: ##EQU10## where C i is the ith neuron's input capacitance; R i is the ith neuron's resistance, T ij is the connection strength between the ith and jth neurons, f(s) is a smooth, monotonically nondecreasing function, and I i is a bias current, comprising: N═(p+r) M neurons, each neuron comprising an analog amplifier, a distinct neuron being assigned to represent the value of each state variable at each sampling time and a distinct neuron being assigned to represent the value of each control variable at each sampling time; an input for each neuron; an output for each neuron; an external input connected to the neurons representing the initial values of each state variable and each control variable, the external input having a value determined by a modified quadratic cost function J': ##EQU11## where R is a p×p state transition matrix, B is a p×r control distribution matrix, x k is a p×1 state vector, u k is an r×1 control vector at the sampling time instant k, Q and R are symmetric weighting matrices, M is the number of time steps, λ is a finite constraint strength, and I is the identity matrix; and a connection between the output of each neuron and the input of each other neuron in the network, the strength of each such connection being determined by the modified cost function, such that running the neural network for a sufficient time to minimize the cost function will produce optimum values for each control variable at each sampling time.

Claim 9

The method of claim 5, wherein each neuron further comprises an analog amplifier.

Claim 10

The method of claim 6, wherein the network further comprises a Hopfield network, where the change with time of the state variables s i in the network is governed by the coupled nonlinear differential equations: ##EQU13## where C i is the ith neuron's input capacitance; R i is the ith neuron's resistance, T ij is the connection strength between the ith and jth neurons, f(s) is a smooth, monotonically nondecreasing function, and I i is a bias current.