This work uses fitness sharing in genetic programming to preserve population diversity. As a benchmark, the system has been applied to learning Boolean definitions for multiplexer circuits from data consisting of the input/output relationships of the circuits. In learning two-address, four data line multiplexers, the system exhibits an order of magnitude improvement in error rate on the standard genetic programming approach. In learning three-address, eight data line multiplexers (in a setting which also requires the system to show some ability to generalise), the system exhibits better than a two-to-one reduction in error rate.