An algorithm is proposed for determining the set of corresponding to the COLREG variants of combined action (alteration course with the speed) to avoid collision with several vessels. When establishing compliance with COLREG, 11 types of meeting of the dangerous target and own ship and four stages (planning of maneuver, early, possible and urgent measures) at the segment of COLREG performance in relation to this target were distinguished. It was assumed that the action to prevent a collision is determined at the first stage. We used the general and particular indicators of the COLREG compliance for each of possible variants of combined action. Due to the large number of such options, the formulas used in their analysis for forecasting, assessment of safety and other goals were simplified so that the calculation time was acceptable. On this basis, the danger domain for the targets was selected simple, circular in shape, with the center shifted relative to the target to take into account the greater risk of crossing its course along the bow than at the stern. It was supposed acceptable to replace the trajectories of combined actions with a combination of a straight segment and an arc of a circle; to represent the change in speed during braking by a second-order polynomial; to consider as unrelated the course and speed changes in the joint operation. An algorithm has been developed for obtaining a matrix of general COLREG compliance estimates of possible combined action variants and flow chart of this algorithm is given. By the elements of this matrix, it is possible to search for an effective variant of the passing by under various assumptions and criteria. As an example, an algorithm for calculating the combined action to avoid a collision with a minimum reduction of the speed is given. The reliability of the proposed algorithms was confirmed by simulation modeling of ships collision avoiding processes in various situations. Key words: collision avoidance, combined action, compliance with COLREG, collision avoidance algorithm.