Bayesian optimization error: `ValueError: (...) is not in list`

[stijnh]

Sometimes, Bayesian optimization randomly gives the following exception. I've been trying to make a reproducible example, but have not been succesfull so far (maybe a race condition)? I will do some digging.

  File "/home/sheldens/miniconda3/lib/python3.13/site-packages/kernel_tuner/interface.py", line 696, in tune_kernel
    results = strategy.tune(searchspace, runner, tuning_options)
  File "/home/sheldens/miniconda3/lib/python3.13/site-packages/kernel_tuner/strategies/bayes_opt.py", line 147, in tune
    bo = BayesianOptimization(
        parameter_space,
    ...<5 lines>...
        cost_func,
    )
  File "/home/sheldens/miniconda3/lib/python3.13/site-packages/kernel_tuner/strategies/bayes_opt.py", line 275, in __init__
    self.initial_sample()
    ~~~~~~~~~~~~~~~~~~~^^
  File "/home/sheldens/miniconda3/lib/python3.13/site-packages/kernel_tuner/strategies/bayes_opt.py", line 530, in initial_sample
    self.update_after_evaluation(observation, index, params)
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/sheldens/miniconda3/lib/python3.13/site-packages/kernel_tuner/strategies/bayes_opt.py", line 435, in update_after_evaluation
    del self.unvisited_cache[self.find_param_config_unvisited_index(param_config)]
                             ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^
  File "/home/sheldens/miniconda3/lib/python3.13/site-packages/kernel_tuner/strategies/bayes_opt.py", line 398, in find_param_config_unvisited_index
    return self.unvisited_cache.index(param_config)
           ~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^
ValueError: (np.float64(0.30000000000000004), np.float64(0.30000000000000004), np.float64(0.30000000000000004), np.float64(0.1), np.float64(0.30000000000000004), np.float64(0.30000000000000004), np.float64(0.5)) is not in list

[stijnh]

It seems that, for small search spaces, the LHS can draw the same sample multiple times when taking the initial samples.

The fix would be to make sure the output of the LHS contains no duplicates.

[stijnh]

Fixed in #336