We present a distributed algorithm for jointly adapting transmit powers, beams,
and linear receiver filters in a TDD MIMO interference network. Neither the
transmitters nor receivers have a priori Channel State Information. The
algorithm is based on forward-backward training in which pilot symbols are
alternately transmitted in the forward (transmitters to receivers) and reverse
directions. The beams and receiver filters are updated directly as adaptive
filters and the powers are updated  by an ``analog'' version of interference
pricing: Forward training is used to estimate the interference prices, which are
synchronously transmitted in the backward direction in analog form to estimate
an interference cost. Numerical results are presented that illustrate the
performance of the method.