Servo - September 2007

DIFFERENT BITS

see the following behavior from your circuit:

1) The LEDS perform their startup blink and the “hi!” message prints out the serial port.

2) Initial weight values print out the serial port, I/O lights blink and the top LED on the error meter lights up.

3) Every 50 iterations, a count and error measurement print out the serial port.

FIGURE 7. Input/ output display.

FIGURE 6. Error level meter.

4) The error level starts to fall lower; the level meter changes accordingly.

• feedforward receives the current input values and calculates each neuron’s current activation.

• train receives the current input values and the desired output value. It calls feedforward to activate the network and get the output_activation. It then measures the actual output against the desired output and backpropagates the error, adjusting each weight value accordingly.

tests to ensure that the serial port and LEDs are working. It initializes random hidden and output weights and iterates through a training loop 8,000 times or until the root mean square error is less than 0.35 (it usually takes about 400 iterations). Every 50 iterations, it prints the count and current error value. When the loop completes, it prints how many iterations it took, the final error value, and loops once through an LED display sequence demonstrating what it has learned.

5) After 400-8,000 iterations, the network has learned the XOR function (have patience!). The last LED in the error level meter lights up and the program performs its ending dance.

Final Thoughts

• main initializes the microcontroller settings and performs a couple quick

After programming, you should

This is a simple circuit to display neural network principles, but the same ideas and basic code apply to more complex problem sets. Experiment with the code and circuit and try to create your own input/ output mappings. Think about what you would like to teach a robot to learn, and how you could modify this code so that your circuit could learn from experience and feedback from its environment. Have fun! SV