We envision a humanoid robot to serve as a source of additional motion-support forces in assistance for frail persons. In this context, we present a control strategy for a humanoid to adaptively regulate its assistive force contribution. First, we identify a human model torque control for optimal execution of a priori known motion task from sample recordings of this task performed by a healthy individual. We utilize the identified model in the proposed position discrepancy based observer of the human torque contribution, the unknown and unmeasurable variable. We propose an experience-based human contribution model learning strategy that allows to improve the human contribution estimate from trial-to-trial. The target assistive torque contribution is then calculated as the difference between the optimal torque required for the motion task and the estimated human contribution. The target assistive torque is integrated into a multi-robot quadratic programming task-space controller to compute the desired interaction force required for the robot to supply the necessary assistive torque for the human model. We use the non-optimal recordings of the human motion task to emulate human frailty and apply our adaptive force control strategy to demonstrate the results of a humanoid successfully assisting the simulated human model to restore the optimal motion task performance.