Abstract. Optimal control problems of switching type with linear state dynamics are ubiquitous in applications of stochastic optimization. For high-dimensional problems of this type, solutions which utilize some convexity related properties are useful. For such problems, we present novel algorithmic solutions which require minimal assumptions while demonstrating remarkable computational efficiency. Furthermore, we devise procedures of the primal-dual kind to assess the distance to optimality of these approximate solutions.Key words. optimal control, American options, pathwise stochastic control, duality DOI. 10.1137/S0040585X97T9879101. Introduction. When making decisions under uncertainty, the major difficulty is to determine how to update estimates and decisions in order to achieve optimality over a given time period. These kinds of questions are often framed within the realm of Markov decision theory, which can be viewed as discrete-time optimal stochastic control.The theoretical underpinnings of Markov decision theory are now well understood. Rigorous mathematical treatments are available in textbook form (see [2], [4], [12], and [23]). However, practical applications remain persistently challenging despite the rich arsenal of theoretical tools that are currently available. In this context, approximate dynamic programming (see [22]) grew from attempts to provide simultaneously practically implementable heuristics and theoretical insights as to why they perform well in practice.In order to control a large system, a practical approach to dealing with the high dimensionality of the state space is to first achieve a finite discretization of it. Alternatively, one can rely on an efficient approximation of functions on this space. In this spirit, function-based methods suggest approximating value functions on the state space. One such method is the least squares Monte Carlo approach, which suggests an approximation by a suitably parameterized set of basis functions. As these parameters are computed by performing successive regressions, this method is placed within the regression-based method family.Following [7], [28], [29], the contribution [18] became the source of subsequent research focused on its theoretical justification. Convergence issues are addressed in [8] and later generalized in [27], [9], and [10], and extensions to multiple exercise rights were considered in [6] and studied in [3], where the connection to statistical learning theory and the theory of empirical processes is emphasized. For an overview of the applications of Monte Carlo methods in financial engineering, we refer the interested reader to Glasserman's book [13] and to the literature cited therein. Beyond financial applications, function approximation methods have also been used to capture local behavior of value functions, and advanced regression methods; e.g., kernel methods [20],