A three-layer artificial neural network was used for adaptive control of gait swing generated by neuromuscular electrical stimulation (NMES) in a spinal cord injured subject. Network inputs consisted of knee and ankle goniometer signals for System 1, and knee and hip angular data for System 2. Controller output was proportional to changes in applied NMES pulse width (PW). Stimulation was applied to the left femoral and common peroneal nerves. The neural networks were trained off-line and on-line. Network performance was assessed by applying a number of different stimulation PWs and later comparing the resulting motion to a sample good step observed during the same test session. On-line training consisted of negative and positive reinforcement applied at chosen times. Both on-line and off-line training algorithms consisted of an enhanced supervised backpropagation scheme. Performance evaluation results favour the use of System 1 over System 2. Also, a network trained off-line and later submitted to on-line punishment appears to be more reliable (in automatic mode) than the same network after it is submitted to on-line reward or to off-line training alone. Finally, the systems' immediate response to on-line learning was favourable in all cases.Based on the results, a version of System 1 was used to generate walking in the test subject. This test indicated that the system is promising.