Function Figure 1- A single node in a neural network Input Connections

A neural network is useful for generating a solution to problems without having to build underlying principles into the solution beforehand. Here, the pole-balancing problem is studied: a pole begins in a nearly upright position, standing on one end. At each time step, a neural network, which takes as inputs the position, velocity, angular position and angular velocity, must generate a move that keeps the pole upright. It was found that a network which judges error based on deviation from vertical can keep the pole pennanently upright relatively quickly. Background Simple computing elements-The units of a network A neural network is called that because it is based on a model of the brain. In the brain, neurons have two states, an excited, or fIring state, and a resting state. Dendrites act as inputs, receiving signals from other neurons, while the axon acts a transmitter, sending out an excitation signal. If enough excitation signals are received from the dendrite, the entire neuron goes into an excited state, and the an electric pulse travels along the axon, where it will meet other dendrites. Each node in a neural network is similar to the neuron described above (see figure I). Each unit has several input connections Aj and uses an file:///C|/DOCUME~1/selman/LOCALS~1/Temp/~LWF0003.htm (1 of 4) [12/6/2001 8:27:15 PM] ~LWF0003 internal function 9 (called the activation function) to send an output signal Ai to other nodes. The activation function 9 takes as input the sum of th