We present algorithms for inferring the cost function and reference trajectory from human demonstrations of hand-writing tasks. These two key elements are then used, through optimal control, to generate an impedance-based controller for a robotic hand . The key novelty lies in the flexibility of the feature design in the composition of the cost function, in contrast to the traditional approaches that consider linearly combined features. Cross-entropy-based methods form the core of our learning technique, resulting in sample-based stochastic algorithms for task encoding and decoding. The algorithms are validated using an anthropomorphic robot hand. We assess that the correct compliance is well encapsulated by subjecting the robot to perturbations during task reproduction.