1

I am trying to use simulated annealing / basinhopping to find a minimum description length of a grammar object and the data according to that grammar.

I define the grammar object myself, as well as a take_step (which is a mutation on the grammar):

def total_energy(self, grammar_and_data):  #the objective function

    # ...operations....

    return encoder.totalLength(grammar,data)

def optimize(self, data):

    initial_grammar_and_data = array((initial_grammar, data))
    take_mutation = MyTakeStep()

    ret = basinhopping(self.total_energy,
                 initial_grammar_and_data,
                 niter=100,
                 T=1.0,
                 #stepsize=0.5,
                 #minimizer_kwargs=None,
                 take_step=take_mutation,
                 #accept_test=None,
                 #callback=None,
                 #interval=50,
                 #disp=False,
                 #niter_success=None
                 )

The problem: I get an error:

File "C:\Users\Sefi\Anaconda3\lib\site-packages\scipy\optimize\optimize.py", line 556, in _approx_fprime_helper
    grad[k] = (f(*((xk + d,) + args)) - f0) / d[k]
TypeError: unsupported operand type(s) for +: 'Grammar' and 'float'

How can I make the algorithm refer to the take-step function -alone- in order to find neighbors near the current value of the objective function? A small step-size is not definable here.

Perhaps the basin hopping cannot work with non-numerical arguments? If so, are you aware of generalized simulated annealing algorithms in Python that will do the trick (I could write one myself, but it will be far less efficient than scipy...)?

ali_m
  • 71,714
  • 23
  • 223
  • 298
Sefi Potashnik
  • 11
  • 3
  • 1
    Basinhopping is not suitable for discrete optimization since the local minimization step requires that the objective function is continuous. If you want to go down the simulated annealing route you could perhaps try the [`simanneal`](https://github.com/perrygeo/simanneal) package. A [genetic algorithm](http://stackoverflow.com/a/31495631/1461210) might also make sense. It's not really possible to make a meaningful recommendation without knowing a lot more about the specific problem you're trying to solve. – ali_m Oct 19 '15 at 23:11
  • Understood. Thank you for the answer and reference! – Sefi Potashnik Oct 20 '15 at 08:10
  • 2
    What you say is more or less true in practice. But in principle basinhopping could work for discrete optimization as long as you chose a local minimizer that could handle discrete functions. – Jacob Stevenson Dec 04 '15 at 17:43

0 Answers0