Tactic Pruning Method in Mixed-Discrete Optimization

• Tactic pruning method is a targeted rule-based approach that eliminates dominated subproblems in mixed-discrete bi-objective optimization.
• It employs a two-phase process where Phase A uses rigorous utopia point dominance and Phase B applies a heuristic center-point test.
• Benchmark studies show up to 96% reduction in subproblem solves while preserving the complete Pareto front.

A tactic pruning method is a targeted, rule-based approach designed to accelerate the search for optimal or valid solutions in complex, combinatorial, or multi-objective settings by identifying, justifying, and removing large subproblems or decision branches that can be rigorously certified as unable to contribute to the optimal solution set. The “tactic pruning method” is particularly influential in mixed-discrete bi-objective optimization, as exemplified by the two-phase framework of Hong, Ahn, and Choi for global Pareto front generation (Hong et al., 2013).

1. Mixed–Discrete Bi-Objective Problem Landscape

Mixed-discrete bi-objective optimization requires finding the Pareto-optimal set of solutions $X^*$ defined by:

$$X^* = \{x\in X \mid \nexists x'\in X,\,J(x')\leq J(x),\,J(x')\neq J(x)\},$$

where $x = [y; z]$ consists of continuous decision variables $y\in\mathbb{R}^{n_y}$ (with box constraints $l_i\leq y_i\leq u_i$) and discrete variables $z\in Z_1\times\cdots\times Z_{n_z}$, with finitely enumerated $Z_j$ sets.

The optimization is formalized as: $\min J(x) = [J_1(y, z), J_2(y, z)]^T,$ subject to

$$g(y, z)\leq 0, \quad h(y, z) = 0, \quad y\in[l, u].$$

The challenge is the combinatorial explosion of discrete assignments, each requiring a continuous bi-objective subproblem solve. Tactic pruning methods allow aggressive, mathematically grounded elimination of many subproblems before full exploration.

2. Algorithmic Decomposition: Continuous Subproblems

The problem is decomposed into $K = \prod_j |Z_j|$ continuous subproblems $P_k$: $$(P_k): \min J([y, z_k]) \text{ s.t. } g(y, z_k)\leq 0,\,h(y, z_k)=0,\,y\in[l, u].$$ Each $z_k$ defines $Y_k$, its feasible $y$-set. The Pareto set $X^*_k$ for each $k$ is the set of non-dominated solutions within that discrete context.

A naive union of all $X^*_k$ (over $k$) is always correct but computationally prohibitive. Tactic pruning seeks to determine, for each $k$, whether $P_k$ can be rigorously excluded from further consideration before costly Pareto front computation.

3. Phase A Pruning — Utopia-Point Dominance

Phase A targets subproblem elimination using the geometry of "anchor" and "utopia" points.

Step A-1: Anchor and Utopia Point Construction

For each discrete candidate $z_k$, compute:

• Anchor points:
  • $y_{k,1}^a = \arg\min_{y\in Y_k} J_1(y, z_k)$,
  • $y_{k,2}^a = \arg\min_{y\in Y_k} J_2(y, z_k)$.
• Anchor objectives:
  • $J_{k,1}^a = J([y_{k,1}^a, z_k])$,
  • $J_{k,2}^a = J([y_{k,2}^a, z_k])$.
• Utopia point:
  • $U_k = [J_{k,1}^{a,1},\,J_{k,2}^{a,2}]^T$.

Step A-2: Master Front via Utopia-Point Pareto Filtering

• Identify the set $K_1^m$ of subproblems whose utopia points are not dominated by any other utopia point: $$K_1^m = \{k \mid \nexists\,\ell\neq k,\,U_\ell \leq U_k\}$$.
• Compute the full Pareto front $X^*_k$ for each $k\in K_1^m$.
• The master front $M$ is defined as $M = \bigcup_{k\in K_1^m} X^*_k$.

Step A-3: Prune Subproblems via Utopia Point Dominance Over $M$

• For $k\notin K_1^m$, if there exists an $x\in M$ such that $U_k \geq J(x)$, then any solution from $P_k$ would be dominated globally—these $k$ indices are pruned.
• Define $$K_0^u = \{k\notin K_1^m \mid \exists x\in M,\,U_k\geq J(x)\}$$.
• Surviving subproblems: $K_1^u = \{1,\dots,K\}\setminus K_0^u$.

Mathematical guarantee: Every point from $P_k$ for $k\in K_0^u$ is dominated by at least one point in $M$; thus, Phase A is lossless and safe (Hong et al., 2013).

4. Phase B Pruning — Center-Point Dominance (Heuristic)

Phase B is a heuristic intended to make additional eliminations if Phase A leaves too many subproblems.

• For $k\in K_1^u\setminus K_1^m$, compute a single "center" objective point (e.g., with weight $\lambda=0.5$):

$$y_k^c = \arg\min_{y\in Y_k} 0.5\,J_1(y, z_k) + 0.5\,J_2(y, z_k),\quad C_k = J([y_k^c, z_k]).$$

• Prune $k$ if $C_k$ is dominated by any $J(x)$ with $x\in M$. Those $k$ form $K_0^c$.
• The survivors $K_1^c = K_1^u\setminus K_0^c$ are fully solved for their Pareto fronts.
• Phase B may potentially prune valid (but marginal) contributors to the global front, so is optional if strict Pareto completeness is required.

5. Integrated Workflow and Complexity Analysis

The algorithmic pipeline is:

1. Enumerate all discrete $z_k$.
2. For all $k$, compute anchor and utopia points.
3. Identify nondominated utopia points, forming $K_1^m$.
4. Solve full Pareto fronts for all $k\in K_1^m$ to construct $M$.
5. Apply Phase A pruning, then (optionally) Phase B pruning.
6. Compute remaining Pareto fronts and assemble the global nondominated union.

Complexity reduction: The pruning method achieves an order-of-magnitude reduction in nonlinear programming (NLP) solves versus brute-force. E.g., in the Van Veldhuizen test ($K=121$), only 4 subproblems of 121 need full solution for the global front after pruning—>96% pruning ratio (Hong et al., 2013).

Benchmark	Total Subproblems ($K$)	After Phase A	After Phase B	True Pareto Discrete Cases
Van Veldhuizen	121	5	4	4
Nine-Bar Truss	4096	~900	72	72

All Pareto fronts obtained are identical to what exhaustive mixed-integer optimization would produce, but with dramatically fewer subproblem solves.

6. Generalizations and Limitations

Phase A generalizes to $n$ objectives by replacing the utopia point dominance test with higher-dimensional bounding constructs (e.g., convex hulls of individualized minima), with attendant increases in combinatorial bookkeeping. If the discrete variable space is extremely large, even anchor point enumeration may become infeasible and alternative meta-heuristic or branch-and-bound sampling approaches are required.

The safe pruning (Phase A) ensures no Pareto-optimal solution is lost. The Phase B heuristic, while effective in practice, is not strictly guaranteed to preserve all global optima but often yields strikingly efficient pruning ratios.

7. Significance and Extensions Across Disciplines

The tactic pruning method offers a canonical model for how proof-theoretic, geometric, or combinatorial certificates (here, utopia point analysis) can enable early rejection of massive swathes of search space across optimization, combinatorial problem solving, and decision science. In mixed-integer nonlinear programming and multi-objective engineering design, the approach exemplifies rigorous integration of problem substructure into global solution strategies (Hong et al., 2013). Its foundational elements—anchoring, global domination tests, Pareto filtering—recur in advanced pruning techniques for other hybrid discrete-continuous or multi-objective settings.