Results of the GECCO'2016 3-OBJ Track
"Raw" Result Data
On each problem participants were judged by the overall dominated hypervolume within the given budget of function evaluations. There were 7 participants in the field. The best function value per problem and participant (1000 times 7 double precision numbers) is listed in this text file.
Participants were ranked based on aggregated problem-wise ranks (details here and here). The following results table lists participants with overall scores (higher is better) and the sum of ranks over all problems (lower is better) The table can be sorted w.r.t. these criteria.
|rank||participant||method name||method description||software||paper||score||sum of ranks|
|1||DIKU, University of Copenhagen||General UB-MO-CMA-ES||UB-MO-CMA-ES generalisation for more than 2 objectives. No recombination of matrices and the largest contributor is selected instead of being chosen proportional to its contribution||956.985||2034|
|2||Simon Wessing||Restarted local search + SMS-EMOA||
This is a rather ad-hoc approach, recycling the existing code for the single-objective tracks. The first 75% of the budget are spent on single-objective restarted local search. An objective of the problem is randomly chosen and a local search is run on it. Then, another objective is randomly chosen and the previously found optimum is used as starting point for this objective. This strategy is iterated until the allocated budget is exhausted. The motivation is to find some hopefully Pareto-optimal local optima, if the problem is multimodal. The last 25% of the budget are spent on a (100 + 5)-SMS-EMOA with self-adaptive Gaussian mutation and no recombination, started from the non-dominated points found so far. This algorithm explicitly optimizes hypervolume and thus is hopefully able to provide some refinement.
|4||Mohammadamin Jahanpour and Bryan Tolson||Pareto archived dynamically dimensioned search (PA-DDS)||Convex Hull Contribution was used as selection metric for PA-DDS||link||link||185.18||3987|
|5||Artelys||Artelys Knitro||Artelys Knitro used in derivative-free mode with multistart||link||link||152.083||4920|
|7||Al Jimenez||Curved Trajectories Algorithm (CTA)||link||31.2688||6219|
Visualization of Performance Data
The following figure shows an aggregated view on the performance data.
The following figures show the same data, but separately for each problem dimension.