1: \begin{abstract}

2: This paper proposes a constrained stochastic successive convex approximation

3: (CSSCA) algorithm to find a stationary point for a general non-convex

4: stochastic optimization problem, whose objective and constraint functions

5: are non-convex and involve expectations over random states. Most existing

6: methods for non-convex stochastic optimization, such as the stochastic

7: (average) gradient and stochastic majorization-minimization, only

8: consider minimizing a stochastic non-convex objective over a deterministic

9: convex set. The proposed CSSCA algorithm can also handle stochastic

10: non-convex constraints in optimization problems, and it opens the

11: way to solving more challenging optimization problems that occur in

12: many applications. The algorithm is based on solving a sequence of

13: convex objective/feasibility optimization problems obtained by replacing

14: the objective/constraint functions in the original problems with some

15: convex surrogate functions. The CSSCA algorithm allows a wide class

16: of surrogate functions and thus provides many freedoms to design good

17: surrogate functions for specific applications. Moreover, it also facilitates

18: parallel implementation for solving large scale stochastic optimization

19: problems, which arise naturally in today's signal processing such

20: as machine learning and big data analysis. We establish the convergence

21: of CSSCA algorithm with a feasible initial point, and customize the

22: algorithmic framework to solve several important application problems.

23: Simulations show that the CSSCA algorithm can achieve superior performance

24: over existing solutions.

25: \end{abstract}