Decomp_cholesky bug

[vlall]

This error terminates the leaner and happens on ~20% of runs for active learner and 60-70% of the time on the BALD learner.. The higher the budget, the more likely you encounter the error. It usually occurs in the 46th+ iteration of training, but sometimes occurs as soon as iteration 12.

My config was:

POOL_SIZE = 500
BUDGET = 50
BASE_KERNELS = ["PER", "LIN", "K"]
DEPTH = 2

This occurs with RandomStrategy and BALD

  File "learn_from_trial.py", line 47, in <module>
    learner.update(correct, guess)
  File "/Users/vlall/prog/dev/bams/bams/learners.py", line 60, in update
    self.models.update()
  File "/Users/vlall/prog/dev/bams/bams/models.py", line 142, in update
    model.update()
  File "/Users/vlall/prog/dev/bams/bams/models.py", line 35, in update
    self.gp.compute(self.data.x, yerr=self.yerr)
  File "/Users/vlall/.virtualenvs/bams/lib/python2.7/site-packages/george/gp.py", line 309, in compute
    self.solver.compute(self._x, yerr, **kwargs)
  File "/Users/vlall/.virtualenvs/bams/lib/python2.7/site-packages/george/solvers/basic.py", line 68, in compute
    self._factor = (cholesky(K, overwrite_a=True, lower=False), False)
  File "/Users/vlall/.virtualenvs/bams/lib/python2.7/site-packages/scipy/linalg/decomp_cholesky.py", line 91, in cholesky
    check_finite=check_finite)
  File "/Users/vlall/.virtualenvs/bams/lib/python2.7/site-packages/scipy/linalg/decomp_cholesky.py", line 40, in _cholesky
    "definite" % info)
numpy.linalg.linalg.LinAlgError: 2-th leading minor of the array is not positive definite```

Mac 10.14
Python 2.7.10

[vlall]

Changing the BIC to true avoid this bit of code, which is the purpose of the branch here: https://github.com/Dallinger/bams/tree/bic-true.

See:

bams/bams/models.py

Line 80 in 2010737

def log_evidence(self, bic=False):