ABSTRACT
Adaptive logic networks (ALNs) were used to control a nonlinear mechanical model of a vehicle active suspension system. An ALN consists of a tree of logic gates having linear threshold units (simple perceptrons) at its leaves. ALNs learned to predict future states based on relationships among values of variables recorded during operation. Piecewise linear functions were extracted from trained ALNs, and executed with the aid of a decision tree so that only a small number of linear pieces had to be evaluated to compute any output value. A 486DX2-66 PC was able to produce a control output in 250 /xs, much faster than was required to control the test system in real time. The results are applicable to a broad range of real-time control problems.
