In this paper, by using a biomechanical model of the human body, we prove that (1) due to the existence of bi-articular muscles and compliant-elements, blind full-torque-compensation at joint level leads to muscles' activity amplification and consequently online adaptation methods are required for exoskeleton torque optimization. Moreover, (2) we state a new hypothesis that "reducing the net torque of two antagonistic mono-articular muscles is sufficient for involved muscles' total effort reduction" and analytically discuss its validity condition. Using this hypothesis, (3) we develop an adaptation rule which optimizes the exoskeleton torque using EMG signals of only two antagonistic mono-articular muscles. Furthermore, (4) the stability, convergence, optimality, and robustness of our adaptation method are proved in the presence of electromyography's intrinsic noisy behavior. Finally, (5) we experimentally validate our EMG-based adaptation method on six healthy subjects. We show that adaptation of the elbow compliance in a 2-DOF semi-active assistive arm in a cyclic task results in significant muscles activity reduction in all our subjects.