Tremor is an involuntary movement characterized by regular or irregular oscillations of one or several body segments. Physiological and pathological tremor in motor control can be defined as roughly sinusoidal movements with particular amplitude and frequency profiles. The electrophysiological analysis of human tremor has a long tradition. Tremor time series belongs to stochastic signals. This because the mechanism of generating them is so complex and exposed to so many uncontrollable influence that mathematical equations describing them contain random quantities. In this study, we concerned with tremor classification for the purpose of medical diagnosis. Accelerometer based tremor signals belong to Parkinsonian, essential, and healthy subjects were considered for this aim. Following features were extracted from tremor signals for classification by artificial neural network (ANN); linear prediction coefficients, wavelet transform detail coefficients, wavelet transform based entropy and variance, power ratio, and higher-order cumulants. Scaled-conjugate (SCG) and BFGS (Broyden–Fletcher–Goldfarb–Shanno) gradient learning algorithms were used. Despite BFGS algorithm had more sensitivity value (92.27%), SCG algorithm had more specificity value (89.01%). According to overall performance, BFGS algorithm (91.02%) was better than SCG algorithm (88.48%).