[論文レビュー] Enterprise Resource Planning Using Multi-type Transformers in Ferro-Titanium Industry

Samira Yazdanpourmoghadam, Mahan Balal Pour+1 more|arXiv (Cornell University)|Jan 28, 2026

Scheduling and Optimization Algorithms被引用数 0

ひとこと要約

The paper trains a Multi-Type Transformer (MTT) to solve Knapsack and Job-Shop Scheduling problems in ERP, benchmarks it against an OR solver on standard instances, and validates it on a Ferro-Titanium industry application, showing competitive gap and fast inference on GPUs.

ABSTRACT

Combinatorial optimization problems such as the Job-Shop Scheduling Problem (JSP) and Knapsack Problem (KP) are fundamental challenges in operations research, logistics, and eterprise resource planning (ERP). These problems often require sophisticated algorithms to achieve near-optimal solutions within practical time constraints. Recent advances in deep learning have introduced transformer-based architectures as promising alternatives to traditional heuristics and metaheuristics. We leverage the Multi-Type Transformer (MTT) architecture to address these benchmarks in a unified framework. We present an extensive experimental evaluation across standard benchmark datasets for JSP and KP, demonstrating that MTT achieves competitive performance on different size of these benchmark problems. We showcase the potential of multi-type attention on a real application in Ferro-Titanium industry. To the best of our knowledge, we are the first to apply multi-type transformers in real manufacturing.

研究の動機と目的

• Motivate the use of neural combinatorial optimization for ERP-related NP-hard problems like KP and JSP.
• Propose a unified graph-based MTT framework with type-specific attention to handle heterogeneous problem structures.
• Evaluate MTT on standard KPI benchmarks and a real Ferro-Titanium industry application.
• Demonstrate practical solution quality and speed relative to exact solvers and discuss generalization potential.

提案手法

• Represent KP and JSP as a heterogeneous graph with type-specific attention in the MTT backbone.
• Use the Multi-Type Transformer architecture to learn problem structures and generate high-quality solutions.
• Benchmark MTT against an OR solver on standard KP and JSP instances across multiple sizes.
• Apply a data transformation and cost-reformulation to map the Ferro-Titanium material loading problem to KP.
• Measure optimality gaps and runtimes on GPU (NVIDIA A100) to assess practicality.

Figure 1: Overview of the unified MTT pipeline for ERP optimization, KP and JSP are represented using heterogeneous graphs and solved using a shared Transformer backbone with type-specific attention.

実験結果

リサーチクエスチョン

• RQ1 Can MTT achieve competitive optimality gaps on 0-1 Knapsack and Job-Shop Scheduling across increasing instance sizes?
• RQ2 Does a unified heterogeneous-graph transformer generalize across packing and scheduling tasks within ERP contexts?
• RQ3 How does MTT perform on real-world industrial problems compared to traditional solvers in terms of gap and speed?
• RQ4 What data transformations are effective to align industrial objectives (cost minimization) with the MTT maximization framework?

主な発見

Problem	Size	OR Solver	MTT	Optimality Gap	time (s)
KP	50	24040	23996	0.0019	16
KP	60	27343	27292	0.0019	23
KP	70	30074	30041	0.0011	33
KP	80	32458	32411	0.0015	45
KP	90	34723	34676	0.0014	59
KP	100	36531	36490	0.0011	78
JSP	5×5	410	421	0.027	23
JSP	6×6	502	515	0.024	58
JSP	7×7	580	594	0.025	131
JSP	8×8	653	673	0.031	268
JSP	9×9	726	752	0.035	516
JSP	10×10	817	849	0.039	921

• MTT achieves an optimality gap around 0.001 for Knapsack across sizes 50–100 (OR solver baseline).
• MTT achieves an optimality gap around 0.02–0.03 for JSP across sizes 5×5 to 10×10 (relative to the combinatorial solver).
• In the Ferro-Titanium application, MTT maintains a stable optimality gap between 0.025 and 0.029 while delivering near-optimal loading plans in under a second per instance on a GPU.
• The KP application transformed to align with MTT’s maximization objective yields consistent performance as item count increases (N = 50–100) with gaps 0.026–0.029 in Table 2.
• MTT demonstrates fast inference and competitive solution quality, suggesting potential for industrial ERP optimization when integrated with learning-based warm starts and adapters.

Figure 2: Schematic illustration of combinatorial problems studied, the visualization of KP (left panel), and JSP (right panel).