Multi-objective collision search gets three fixes: IMOCCO tops 31 test functions and holds its edge on truss and brake design

Multi-objective collision search gains three layers of improvement, giving the population separate controls for spreading along the front and converging toward it. The earlier MOCCO replaced CCO's scalar-fitness ordering with fast non-dominated sorting, crowding distance, an external archive and polynomial mutation, but its collision weight was fixed, crowding distance resolved poorly on irregular fronts, and the archive only stored solutions passively. Over 31 standard test functions with 30 independent runs, IMOCCO's Friedman mean ranks are 1.42 on IGD, 2.10 on GD, 1.77 on Spacing and 1.65 on HV, for an overall mean rank of 1.73, first among seven algorithms.

The improvements sit at three layers: the search operator separates motion along the front from motion toward it, individual interaction makes collision strength depend on solution quality and objective-space distance, and environmental selection uses one sparsity measure for both archive pruning and collision-center choice. Manifold-feature collision estimates tangent and normal spaces of the Pareto set from a rank-weighted covariance of the elite half and scales the two displacement types anisotropically; mass-restitution collision sets mass from Pareto quality and restitution from objective-space distance; angular-sparsity guidance prunes the archive by a farthest-first rule in normalized objective space and picks collision centers by roulette within an angular neighborhood. An ablation under identical settings compares full IMOCCO with variants lacking S1, S2 or S3 and with all three off; the full version has an overall mean rank of 2.17, removing S3 raises it to 3.73, removing S1 to 2.61, removing S2 to 2.28, and all off to 4.21.

The demonstrated scope is 31 standard test functions plus two constrained engineering problems, a four-bar truss and a disc brake, all implemented in MATLAB with structural analysis by the direct stiffness method on a Core 2 Duo 2.53 GHz machine. The intended reader is a structural or engineering design optimization researcher who needs a set of compromise Pareto solutions in one run. The authors propose extending the method to larger three-dimensional trusses, combining it with surrogate models to cut computational cost, and applying it to more practical engineering structures.

Beyond population and archive sizes, the current implementation fixes several mechanism-specific coefficients, and the authors state that systematic sensitivity analysis of these coefficients remains future work. On some difficult functions other algorithms remain competitive, and the size of the advantage varies by metric. On DTLZ3 every algorithm reports zero HV, so that function offers no discriminating information.