1: \begin{abstract}

2: We study constrained stochastic programs where the decision 

3: vector at each time slot cannot be chosen freely but is tied to 

4:  the realization of an underlying random state vector. The goal is to 

5: minimize a general objective function subject to linear constraints. 

6: A typical scenario where such programs appear is  opportunistic 

7: scheduling over a network of time-varying channels, where the random state 

8: vector is the channel state observed, and the control vector is the transmission 

9: decision which depends on the current channel state.  

10: We consider a primal-dual type Frank-Wolfe algorithm that 

11: has a low complexity update during each slot and that learns to make

12: efficient decisions without prior knowledge of the 

13: probability distribution of the random state vector. 

14: We establish convergence time guarantees for the case of both convex 

15: and non-convex objective functions.  We also emphasize application of the 

16: algorithm to non-convex opportunistic 

17: scheduling and distributed non-convex stochastic optimization over a connected graph. 

18: \end{abstract}

19: