Table 1.

Genetic algorithms with characteristics below were used to optimize the likelihood function in (2) and produce an ensemble of models.

Method	No. of swarms M	Particles per swarm N	Initialization of particles	Number of generations (iter)	Final ${\chi }^2$
DMS-PSO-CLS29	5	4	Sobol	1,000	4,607.65*
DMS-PSO-CLS	20	4	Sobol	1,000	2,773.07
DMS-PSO-CLS	10	4	Sobol	1,000	2,781.44
DMS-PSO-CLS	10	4	Random	1,000	3,703.9
DMS-PSO-CLS	10	4	Random	600	2,743.5
PSO-DLS28	10	4	Random	600	2,933.5
PSO-DLS	5	4	Sobol	1,000	2,708.05
PSO-DLS	20	4	Sobol	1,000	2,797.88
PSO-DLS	10	4	Sobol	1,000	3,436.22
PSO-DLS	10	4	Random	1,000	2,772.33
PSO-DLS	10	4	Random	1,000	2,880.11
PSO-DLS	10	4	Random	1,000	2,941.89
PSO-DLS	5	4	Random	1,000	6,702.86*
PSO-DLS	20	4	Random	1,000	2,768.15

Each run was initialized with $\theta$-parameters either initially positioned on a space-filling Sobol sequence⁴⁰ or randomly within the 35-dimensional parameter space including the illumination parameter $f_{\mathit{IL}}$ (see “Materials and methods”). All genetic algorithms were run for 600–1,000 generations to equilibrate the search for an optimum to Eq. (2).

*These two algorithms had only 20 particles and were eliminated from further consideration.