Genliang Li, Haixing Chen and Yiqing Cao combined covariance-guided search, rank-based learning and a response to stagnation in a new optimization method. It outperformed its BPBO baseline in the reported benchmark comparisons and reduced route costs in an aircraft simulation. Matrix decomposition adds computational work, tying the gains to the size and conditions of the search problem.
Genliang Li, Haixing Chen and Yiqing Cao developed CGPBPBO, an adaptation of the BPBO optimization algorithm for searching complex solution spaces. A memory of elite solutions uses covariance, a measure of how variables change together, to rotate search directions. Rank-based pursuit learning adjusts candidate behavior by performance. When progress stalls, a mutation mechanism combines Gaussian and Cauchy distributions to change the search.[1]
Tests used the CEC2017 and CEC2022 optimization benchmark suites over 30 independent runs. With the deprecated F2 function excluded, CGPBPBO achieved the best reported mean on 17 of the remaining 29 CEC2017 functions and exceeded baseline BPBO on all 29. It exceeded the baseline in CEC2022 at 10 dimensions across all 12 functions. At 20 dimensions, it won on 11 functions out of 12.[1]
An unmanned-aircraft route-planning simulation returned mean path cost of 239.8048 with standard deviation of 26.2686. Relative to baseline BPBO, the researchers reported reductions of 21.74 per cent in path cost and 59.68 per cent in variability. Those results concern simulation rather than physical flight. Matrix decomposition involves additional computation, a consideration the paper retains for high-dimensional problems.[1]
