Commercially available active noise cancellation (ANC) headphones typically use hybrid single-channel filter structures that consist of a feedforward and a feedback filter. Multi-channel feedforward filtering can improve the achievable performance by exploiting spatial characteristics of the incoming noise. This poses the challenge of determining the ideal positions of multiple reference microphones. We can analyze the position of reference microphones either based on a simulation, such as the finite-element method, or by propping a dummy headphone with multiple microphones. In both cases, the positions can be evaluated for a given number of reference microphones either based on the multiple coherence function (MCF) or by evaluating the ANC performance for the optimal feedforward FIR filter.We propose the solution to a least mean squares cost function for multi-channel feedforward FIR filter design for ANC headphones. Based on this cost function, we develop an algorithm that determines the optimal reference microphone positions for an arbitrary but fixed number of microphones. Finally, we compare the proposed algorithm to that of using the MCF as cost function. The evaluation is based on measurements of a headphone that is equipped with 10 additional digital MEMS microphones.