def _partition_space()

in botorch/utils/multi_objective/box_decompositions/box_decomposition.py [0:0]

• 23 lines of code
• 8 McCabe index (conditional complexity)

    def _partition_space(self):
        r"""Partition the non-dominated space into disjoint hypercells.

        This method supports an arbitrary number of outcomes, but is
        less efficient than `partition_space_2d` for the 2-outcome case.
        """
        if len(self.batch_shape) > 0:
            # this could be triggered when m=2 outcomes and
            # BoxDecomposition._partition_space_2d is not overridden.
            raise NotImplementedError(
                "_partition_space does not support batch dimensions."
            )
        # this assumes minimization
        # initialize local upper bounds
        self.register_buffer("_U", self._neg_ref_point.unsqueeze(-2).clone())
        # initialize defining points to be the dummy points \hat{z} that are
        # defined in Sec 2.1 in [Lacour17]_. Note that in [Lacour17]_, outcomes
        # are assumed to be between [0,1], so they used 0 rather than -inf.
        self._Z = torch.zeros(
            1,
            self.num_outcomes,
            self.num_outcomes,
            dtype=self.Y.dtype,
            device=self.Y.device,
        )
        for j in range(self.ref_point.shape[-1]):
            # use ref point for maximization as the ideal point for minimization.
            self._Z[0, j] = float("-inf")
            self._Z[0, j, j] = self._U[0, j]
        # incrementally update local upper bounds and defining points
        # for each new Pareto point
        self._U, self._Z = update_local_upper_bounds_incremental(
            new_pareto_Y=self._neg_pareto_Y,
            U=self._U,
            Z=self._Z,
        )
        self._get_partitioning()