[Feature Request] Optimization over high-dim discrete / combinatorial search spaces #477
Description
Thank you for the great library!
I'm new to Ax and trying it with simple benchmark functions. I found that the EHVI might be less effective than simple Sobol sampling when I test it with the multi-objective knapsack problem. As can be seen in the following plots, EHVI found fewer solutions. In addition, it tended to violate the constraints (objective1 < 479 and objective2 < 545) and tried to maximize the objectives.
I'm not sure, but this is partly because EHVI-based method sample the same points multiple times, and partly because it samples infeasible region. It samples -1 while the range of the parameters is 0 or 1 as follows:
Parameter definition:
parameters = [RangeParameter(f"x{i}", ParameterType.INT, lower=0, upper=1) for i in range(20)]Log message:
[WARNING 01-21 12:06:27] ax.modelbridge.transforms.int_to_float: Unable to round {'x0': 1, 'x1': -1, 'x2': -1, 'x3': 1, 'x4': 2, 'x5': 1, 'x6': 0, 'x7': -1, 'x8': 1, 'x9': 1, 'x10': -1, 'x11': 1, 'x12': 0, 'x13': 1, 'x14': 0, 'x15': 0, 'x16': 0, 'x17': -1, 'x18': 0, 'x19': -1}to meet constraints of SearchSpace(parameters=[RangeParameter(name='x0', parameter_type=INT, range=[0, 1]), RangeParameter(name='x1', parameter_type=INT, range=[0, 1]), RangeParameter(name='x2', parameter_type=INT, range=[0, 1]), RangeParameter(name='x3', parameter_type=INT, range=[0, 1]), RangeParameter(name='x4', parameter_type=INT, range=[0, 1]), RangeParameter(name='x5', parameter_type=INT, range=[0, 1]), RangeParameter(name='x6', parameter_type=INT, range=[0, 1]), RangeParameter(name='x7', parameter_type=INT, range=[0, 1]), RangeParameter(name='x8', parameter_type=INT, range=[0, 1]), RangeParameter(name='x9', parameter_type=INT, range=[0, 1]), RangeParameter(name='x10', parameter_type=INT, range=[0, 1]), RangeParameter(name='x11', parameter_type=INT, range=[0, 1]), RangeParameter(name='x12', parameter_type=INT, range=[0, 1]), RangeParameter(name='x13', parameter_type=INT, range=[0, 1]), RangeParameter(name='x14', parameter_type=INT, range=[0, 1]), RangeParameter(name='x15', parameter_type=INT, range=[0, 1]), RangeParameter(name='x16', parameter_type=INT, range=[0, 1]), RangeParameter(name='x17', parameter_type=INT, range=[0, 1]), RangeParameter(name='x18', parameter_type=INT, range=[0, 1]), RangeParameter(name='x19', parameter_type=INT, range=[0, 1])], parameter_constraints=[ParameterConstraint(47*x0 + 22*x1 + 82*x2 + 19*x3 + 85*x4 + 15*x5 + 89*x6 + 74*x7 + 26*x8 + 11*x9 + 86*x10 + 81*x11 + 16*x12 + 35*x13 + 60*x14 + 30*x15 + 28*x16 + 94*x17 + 21*x18 + 38*x19 <= 479), ParameterConstraint(39*x0 + 24*x1 + 60*x2 + 78*x3 + 97*x4 + 97*x5 + 96*x6 + 23*x7 + 19*x8 + 17*x9 + 73*x10 + 71*x11 + 32*x12 + 67*x13 + 11*x14 + 10*x15 + 70*x16 + 91*x17 + 18*x18 + 98*x19 <= 545)])I'm wondering whether it was expected behavior or I configured the EHVI wrongly. Here's a notebook to reproduce the experiment. Could you give me any comments about it?