Neural network

A method of optimizing a neural network having dynamically varying operating conditions, said method comprising:

• arranging a plurality of neurons in an input layer receiving an input signal from an external system, a plurality of intermediate layers, and an output layer supplying an output signal to said external system,
• connecting each neuron in an intermediate layer and the output layer to a predetermined number of neurons in the preceding layer,
• connecting each neuron in the input layer and an intermediate layer to said predetermined number of neurons in the subsequent layer,
• receiving by each neuron of one of an intermediate layer and the output layer a combined input signal which is the sum of the output signals of the neurons in the preceding layer,
• adapting by each neuron to fire or emit an electric pulse,
• setting the state n(t) of said neuron to "0" when no pulse is emitted and to "1" when a pulse is emitted if one of the input signal and the combined input signal of a neuron exceeds a given controllable threshold T(t) so that if one of the input signal and the combined input signal is greater than the threshold value T(t) the state of said neuron will be n(t+.DELTA.t)=1,
• providing the input connections to a neuron in one of an intermediate layer and the output layer with adaptable strengths q.sub.1, q.sub.2, . . . q.sub.k from k neurons in the preceding layer with states n.sub.1, n.sub.2, . . . n.sub.k so that a combined input signal is q.sub.1, (t)n.sub.1 (t)+q.sub.2 (t)n.sub.2 (t)+. . .+q.sub.k (t)n.sub.k (t),
• providing the output connections from a neuron in the input layer and an intermediate layer with adaptable strengths q.sub.1, q.sub.2, . . . q.sub.k to k neurons in the subsequent layer with states n.sub.1, n.sub.2, . . . n.sub.k,
• evaluating the response of the network wherein the number of firings from the output layer is used for determining the threshold value so that when the number of firings exceeds a certain value the threshold value increases,
• leaving the threshold value signal unchanged if the number of firings has said certain value,
• reducing the threshold value when the number of firings is below said certain value,
• updating the strength of a signal q.sub.i of an output connection of a neuron with state n to a neuron with a state n.sub.i in accordance with q.sub.i -q.sub.i =q.sub.i (t)+r(t)f(q.sub.i (t))n(t)n.sub.i (t+.DELTA.t), where r(t) is a feedback signal, and n(t) and n.sub.i (t+.DELTA.t) are state signals which are responsive to the feedback signal applied to the network from said external system being controlled by the neural network so that if n(t) and n.sub.i (t+.DELTA.t) both assume the value 1 then r(t)f(q.sub.i (t)) is included in the adjustment of the strength of a connection, and
• adapting said neural network to at least one of changing conditions and new conditions in said external system by changes to at least one of said threshold value signal and said adaptable strengths of said connections.

A method by a neural network for self adapting to changing operating conditions of an external system coupled to said neural network, said method includes providing a neural network having a plurality of neurons arranged on an input layer receiving an input signal from the external system, a plurality of intermediate layers, and an output layer supplying an input signal to the external system, connecting such neuron in an intermediate layer and the output layer to a predetermined number of neurons in the preceding layer, connecting each neuron in the input layer and an intermediate layer to said predetermined number of neurons in the subsequent layer, receiving by each neuron in one of an intermediate layer or the output layer a combined input signal which is the sum of the output signals of the neurons in the preceding layer, emitting by each one of said neurons a pulse signal in response to a received signal which exceeds a threshold value for said one of said neurons, and evaluating the response of the network, the improvements comprising the steps of:

• determining the threshold value from the number of firings from the output layer,
• increasing the threshold value signal if the number of firings exceeds a certain value,
• leaving the threshold value signal unchanged if the number of firings equals said certain value,
• reducing the threshold value signal if the number of firings is below said certain value,
• updating the strength of a signal q.sub.i of an output connection of a neuron with state n to a neuron with a state n.sub.i in accordance with q.sub.i .fwdarw.q.sub.i =q.sub.i (t)+r(t)f(q.sub.i (t))n(t)n.sub.i (t+.DELTA.t), with r(t) being a feedback signal and said n(t) and n.sub.i (t+.DELTA.t) being state signals, said state signals being responsive to the feedback signal applied to the network from said external system being controlled by the neural network so that if n(t) and n.sub.i (t+.DELTA.t) both assume the value 1 then r(t)f(q.sub.i (t)) is included in the adjustment of the strength of a connection, and
• adapting by said neural network to said changing operating conditions of said external system by at least one of said steps of reducing and updating.

A method of optimizing a neural network having neurons with a connection from each one of said neurons to at least one other of said neurons, said method comprising:

• adapting each one of said neurons to emit a pulse signal on its said connection in response to receiving a pulse signal form at least one other of said neurons;
• associating each one of said neurons with control parameters in the form of a threshold value and a strength level for said connection from said each one of said neurons;
• emitting said pulse signal by each one of said neurons based on the associated said threshold value and said strength level for said connection from said each one of said neurons;
• providing a network input to said neural network from an external system and at least one of said neurons emitting said pulse signal in response to said network input from said external system;
• providing a network output with at least one of said neurons emitting said pulse signal in response to said neural network processing said network input from said external system;
• and
• updating at least one of said threshold value and said strength level of said connection for at least one of said neurons in response to said network output by said neural network wherein updating the strength of a signal q.sub.i of a connection for an output of a neuron with state n to a neuron with a state n.sub.i in accordance with q.sub.i -q.sub.i =q.sub.i (t)+r(t)f(q.sub.i (t))n(t)n.sub.i (t+.DELTA.t), r(t) being a feedback signal, and said n(t) and n.sub.i (t+.DELTA.t) being state signals, said state signals being responsive to the feedback signal applied to the network from said external system being controlled by the neural network so that if n(t) and n.sub.i (t+.DELTA.t) both assume the value 1, then r(t)f(q.sub.i (t)) is included in the adjustment of the strength of a connection, said step of updating improves optimization of said neural network to perform in response to said external system.