evox.algorithms.mo.nsga3

Module Contents

Classes

NSGA3

An implementation of the tensorized NSGA-III for many-objective optimization problems.

Functions

_get_table_row_inner
_select_from_index_by_min_inner
_get_extreme_inner
_masked_assign	Compile-safe replacement for target[index[mask]] = value.

Data

vmap_get_table_row
vmap_select_from_index_by_min
vmap_get_extreme

API

evox.algorithms.mo.nsga3._get_table_row_inner(bool_ref_candidate: torch.Tensor, upper_bound: torch.Tensor)

evox.algorithms.mo.nsga3.vmap_get_table_row

‘vmap(…)’

evox.algorithms.mo.nsga3._select_from_index_by_min_inner(group_id: torch.Tensor, group_dist: torch.Tensor, idx: torch.Tensor)

evox.algorithms.mo.nsga3.vmap_select_from_index_by_min

‘vmap(…)’

evox.algorithms.mo.nsga3._get_extreme_inner(norm_fit: torch.Tensor, w: torch.Tensor)

evox.algorithms.mo.nsga3.vmap_get_extreme

‘vmap(…)’

evox.algorithms.mo.nsga3._masked_assign(target: torch.Tensor, index: torch.Tensor, mask: torch.Tensor, value: torch.Tensor)

Compile-safe replacement for target[index[mask]] = value.

Boolean (advanced) indexing such as target[index[mask]] = value lowers to aten.nonzero, a dynamic-shape operator that graph-breaks under torch.compile / torch.while_loop (which require fullgraph=True).

This performs the same assignment without any dynamic-shape op: for masked-in positions it writes value; for masked-out positions it writes the entry already stored at the (clamped) index, which is a no-op. Duplicate indices are safe because the same scalar value is written.

class evox.algorithms.mo.nsga3.NSGA3(pop_size: int, n_objs: int, lb: torch.Tensor, ub: torch.Tensor, selection_op: Optional[Callable] = None, mutation_op: Optional[Callable] = None, crossover_op: Optional[Callable] = None, data_type: Optional[torch.dtype] = None, device: torch.device | None = None)

Bases: evox.core.Algorithm

An implementation of the tensorized NSGA-III for many-objective optimization problems.

References:

[1] K. Deb and H. Jain, “An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach, Part I: Solving Problems With Box Constraints,” IEEE Transactions on Evolutionary Computation, vol. 18, no. 4, pp. 577-601, 2014. Available: https://ieeexplore.ieee.org/document/6600851

[2] H. Li, Z. Liang, and R. Cheng, “GPU-accelerated Evolutionary Many-objective Optimization Using Tensorized NSGA-III,” in 2025 IEEE Congress on Evolutionary Computation, 2025.

Initialization

Initializes the NSGA-III algorithm.

Parameters:

• pop_size – The size of the population.

• n_objs – The number of objective functions in the optimization problem.

• lb – The lower bounds for the decision variables (1D tensor).

• ub – The upper bounds for the decision variables (1D tensor).

• selection_op – The selection operation for evolutionary strategy (optional).

• mutation_op – The mutation operation, defaults to polynomial_mutation if not provided (optional).

• crossover_op – The crossover operation, defaults to simulated_binary if not provided (optional).

• data_type – The data type for the decision variables (optional). Defaults to torch.float32.

• device – The device on which computations should run (optional). Defaults to PyTorch’s default device.

init_step()

Perform the initialization step of the workflow.

Calls the init_step of the algorithm if overwritten; otherwise, its step method will be invoked.

step()

Perform the optimization step of the workflow.

_get_extreme(norm_fit: torch.Tensor, w: torch.Tensor)

_compute_distances(fit: torch.Tensor, ref: torch.Tensor)