Characterizing Optimal Mixed Policies: Where to Intervene and What to Observe

Neural Information Processing Systems 

Most policies can be parametrized in terms of these two dimensions, i.e., as a function of what can be seen and done