[en] Purpose: Simulated annealing (SA) allows for the implementation of realistic biological and clinical cost functions into treatment plan optimization. However, a drawback to the clinical implementation of SA optimization is that large numbers of beams appear in the final solution, some with insignificant weights, preventing the delivery of these optimized plans using conventional (limited to a few coplanar beams) radiation therapy. A preliminary study suggested two promising algorithms for restricting the number of beam weights. The purpose of this investigation was to compare these two algorithms using our current SA algorithm with the aim of producing a algorithm to allow clinically useful radiation therapy treatment planning optimization. Method: Our current SA algorithm, Variable Stepsize Generalized Simulated Annealing (VSGSA) was modified with two algorithms to restrict the number of beam weights in the final solution. The first algorithm selected combinations of a fixed number of beams from the complete solution space at each iterative step of the optimization process. The second reduced the allowed number of beams by a factor of two at periodic steps during the optimization process until only the specified number of beams remained. Results of optimization of beam weights and angles using these algorithms were compared using a standard cadre of abdominal cases. The solution space was defined as a set of 36 custom-shaped open and wedged-filtered fields at 10 deg. increments with a target constant target volume margin of 1.2 cm. For each case a clinically-accepted cost function, minimum tumor dose was maximized subject to a set of normal tissue binary dose-volume constraints. For this study, the optimized plan was restricted to four (4) fields suitable for delivery with conventional therapy equipment. Results: The table gives the mean value of the minimum target dose obtained for each algorithm averaged over 5 different runs and the comparable manual treatment plan value. The inter-run variation for the combination algorithm was less than 2 Gy compared to 3.8 Gy for the reduction algorithm. In 5 of 6 cases the combination algorithm gave higher minimum target volume doses. Optimization times averaged 9 minutes compared to over 40 minutes for a manually-produced treatment plan with a reduced cost function. It is interesting to note that SA optimization restricted to 4 beams gives results comparable to earlier optimization work using 36 open field beams. Whether this is due to the np-complexity of the 36 beam solution space or to the optimal nature of the 4 beam solution requires further research. Conclusion: 1) Either algorithm allows for a significant dose escalation compared to manual treatment plans. 2) The beam combination algorithm generally produced higher target volume doses with less variation among optimization runs. 3) Optimization runs take less than 10 minutes using current CPUs compared to 40 minutes for a manual treatment plan. 4) SA optimization runs using 4 beams with wedges give results comparable to previous optimization runs using 36 open-field beams