Tests failing in Julia 1.6 (multithreading and RNGs)

[ablaom]

A test in test/adaptive.jl is failing in Julia 1.6. I suppose this has something to do with changes in Julia around RNGs for multithreaded computations.

The test involves comparison of best score for adaptive PSO against a RandomSearch baseline for different acceleration options. For every option the test passes under Julia 1.4 and indeed the loss and base-line losses are identical for each option (I've added debug lines to print these below):

best_loss = 0.08947526016044464
baseline_best_loss = 0.09009691034711517
Test Summary:                                               | Pass  Total
EvoTree Tuning with AdaptiveParticleSwarm and CPU1{Nothing} |    1      1
best_loss = 0.08947526016044464
baseline_best_loss = 0.09009691034711517
Test Summary:                                                       | Pass  Total
EvoTree Tuning with AdaptiveParticleSwarm and CPUProcesses{Nothing} |    1      1
best_loss = 0.08947526016044464
baseline_best_loss = 0.09009691034711517
Test Summary:                                                     | Pass  Total
EvoTree Tuning with AdaptiveParticleSwarm and CPUThreads{Nothing} |    1      1
    Testing MLJParticleSwarmOptimization tests passed 

However, in Julia 1.6 (and 1.7?) the output is different for multithreads and the test fails:

best_loss = 0.0893142957928955
baseline_best_loss = 0.09001814389260604
Test Summary:                                               | Pass  Total
EvoTree Tuning with AdaptiveParticleSwarm and CPU1{Nothing} |    1      1
best_loss = 0.0893142957928955
baseline_best_loss = 0.09001814389260604
Test Summary:                                                       | Pass  Total
EvoTree Tuning with AdaptiveParticleSwarm and CPUProcesses{Nothing} |    1      1
best_loss = 0.09106885087618996
baseline_best_loss = 0.09001814389260604
EvoTree Tuning with AdaptiveParticleSwarm and CPUThreads{Nothing}: Test Failed at /Users/anthony/GoogleDrive/Julia/MLJ/MLJParticleSwarmOptimization/test/strategies/adaptive.jl:122
  Expression: best_loss < baseline_best_loss || isapprox(best_loss, baseline_best_loss; atol = 0.001)
Stacktrace:
 [1] macro expansion
   @ ~/GoogleDrive/Julia/MLJ/MLJParticleSwarmOptimization/test/strategies/adaptive.jl:122 [inlined]                                                                                  
 [2] macro expansion
   @ /Users/julia/buildbot/worker/package_macos64/build/usr/share/julia/stdlib/v1.6/Test/src/Test.jl:1151 [inlined]
 [3] top-level scope
   @ ~/GoogleDrive/Julia/MLJ/MLJParticleSwarmOptimization/test/strategies/adaptive.jl:80
Test Summary:                                                     | Fail  Total
EvoTree Tuning with AdaptiveParticleSwarm and CPUThreads{Nothing} |    1      1
ERROR: LoadError: LoadError: Some tests did not pass: 0 passed, 1 failed, 0 errored, 0 broken.                           
in expression starting at /Users/anthony/GoogleDrive/Julia/MLJ/MLJParticleSwarmOptimization/test/strategies/adaptive.jl:78
in expression starting at /Users/anthony/GoogleDrive/Julia/MLJ/MLJParticleSwarmOptimization/test/runtests.jl:22
ERROR: Package MLJParticleSwarmOptimization errored during testing

I would say the test itself is flawed, as there is no reason to expect PSO to outperform RandomSearch() in a random case. However, this issue raises some questions that need answering:

1. Is it possible to run adaptive PSO in acceleration=CPUThreads mode reproducibly?
2. If so, is it possible to re-implement adaptive PSO so that the acceleration mode makes no difference to the outcome when a user specifies the RNG?
3. What doc fixes or code changes are needed in response to these questions?