Simulated annealing (SA) is a multivariate combinatorial optimization process that searches the configuration space of possible solutions by a random walk, guided only by the goal of minimization of the objective function. The decision-making capabilities of a fuzzy inference system are applied to guide the SA search, to look for solutions which, in addition to optimizing a plan in dosimetric terms, also present some clinically desirable spatial features. No a priori constraints are placed on the number or position of needles or on the seed loading sequence of individual needles. These additional degrees of freedom are balanced by giving preference to consider plans with seed distributions that are balanced in the right/left and anterior/posterior halves in each axial slice, and with local seed density that is about uniform. Piecewise linear membership functions are constructed to represent these requirements. Before a step in the random search is subject to the SA test, the expert functions representing the spatial seed-distribution requirements are evaluated. Thus, the expert planner’s knowledge enters into the decision as to the “goodness” of a seed configuration regarding the spatial seed-distribution goals. When a step in the random walk yields a seed configuration that is found wanting, a specific number of additional steps in the local neighborhood is attempted until either improvement in the spatial requirements is achieved, or the allowed number of attempts is exhausted. In the latter case, the expert system desists and the unfavorable step is taken, moving on to the simulated annealing test. The number of attempts is determined by the fuzzy logic inference engine and depends on just how badly the expert requirement is not met. The program is interfaced with a commercial treatment planning system (TPS) to import optimized seed plans for isodose display and analysis. Execution in a 1.5 GHz computer is less than a minute, adequate for real-time planning. Results for phantom and real patients are dosimetrically comparable to generic methods such as modified peripheral loading, but in some cases, fewer needles and less radioactivity are required. Sensitivity to simulated random seed placement error is similar to manual plans.