Selected work

Representative Papers

On dynamic multi-agent pathfinding methods: review, simulations and modifications

Jun 02, 2026

This work addresses the challenges of dynamic obstacles, partial observability, and agent coordination in dynamic multi-agent pathfinding (D-MAPF) by proposing the A** algorithm. A** introduces a novel template mechanism that decouples offline geometric path generation from online spatiotemporal replanning. By precomputing a diverse set of candidate paths and dynamically reconnecting them during execution, A** efficiently handles environmental changes and perception limitations within a unified simulation framework. The authors evaluate A** against six baseline algorithms—including Dijkstra, D* Lite, Space-Time A*, WHCA*, and M*—demonstrating that A** significantly improves path quality and adaptability in dynamic, partially observable scenarios, thereby validating its effectiveness in complex multi-agent systems.

0 citationsRead paper

Cluster Workload Allocation: Semantic Soft Affinity Using Natural Language Processing

Jan 14, 2026

This work proposes a semantic, intent-driven scheduling paradigm to address the limitations of existing cluster schedulers, which rely on complex configurations and struggle to express soft affinity preferences, leading to a usability gap. For the first time, large language models—such as Amazon Nova Pro/Premier and Mistral Pixtral Large—are integrated into a Kubernetes scheduler extension to interpret user scheduling intents expressed in natural language, enabling semantic soft affinity scheduling. The system leverages a cluster state cache and an AWS Bedrock–based intent analyzer. In evaluations, it achieves over 95% accuracy and matches or outperforms standard Kubernetes configurations across six scenarios, demonstrating particularly strong performance in complex, quantitative, and conflicting preference settings.

0 citationsRead paper

Enhancing Cluster Scheduling in HPC: A Continuous Transfer Learning for Real-Time Optimization

Sep 22, 2025

To address the insufficient real-time adaptability of task scheduling under node affinity constraints in HPC clusters, this paper proposes a dynamic scheduling method based on continual transfer learning. Unlike traditional schedulers (e.g., Kubernetes) that require frequent offline retraining, our approach enables lightweight, online model evolution using incoming task streams—eliminating the need for periodic retraining and significantly reducing operational overhead. Empirical evaluation on the Google Cluster Trace demonstrates that the method maintains over 99% scheduling prediction accuracy while reducing average scheduling latency for affinity-constrained tasks and lowering overall system computational cost. The core contribution is the integration of a continual transfer learning mechanism into the scheduling decision loop, establishing the first affinity-aware, low-overhead, and high-temporal-fidelity adaptive scheduling framework. This advances scalability and dynamic responsiveness in large-scale HPC systems.

0 citationsRead paper

Recent publications

Latest Papers

On dynamic multi-agent pathfinding methods: review, simulations and modifications

Jun 02, 2026

This work addresses the challenges of dynamic obstacles, partial observability, and agent coordination in dynamic multi-agent pathfinding (D-MAPF) by proposing the A** algorithm. A** introduces a novel template mechanism that decouples offline geometric path generation from online spatiotemporal replanning. By precomputing a diverse set of candidate paths and dynamically reconnecting them during execution, A** efficiently handles environmental changes and perception limitations within a unified simulation framework. The authors evaluate A** against six baseline algorithms—including Dijkstra, D* Lite, Space-Time A*, WHCA*, and M*—demonstrating that A** significantly improves path quality and adaptability in dynamic, partially observable scenarios, thereby validating its effectiveness in complex multi-agent systems.

0 citationsRead paper

Cluster Workload Allocation: Semantic Soft Affinity Using Natural Language Processing

Jan 14, 2026

This work proposes a semantic, intent-driven scheduling paradigm to address the limitations of existing cluster schedulers, which rely on complex configurations and struggle to express soft affinity preferences, leading to a usability gap. For the first time, large language models—such as Amazon Nova Pro/Premier and Mistral Pixtral Large—are integrated into a Kubernetes scheduler extension to interpret user scheduling intents expressed in natural language, enabling semantic soft affinity scheduling. The system leverages a cluster state cache and an AWS Bedrock–based intent analyzer. In evaluations, it achieves over 95% accuracy and matches or outperforms standard Kubernetes configurations across six scenarios, demonstrating particularly strong performance in complex, quantitative, and conflicting preference settings.

0 citationsRead paper

Enhancing Cluster Scheduling in HPC: A Continuous Transfer Learning for Real-Time Optimization

Sep 22, 2025

To address the insufficient real-time adaptability of task scheduling under node affinity constraints in HPC clusters, this paper proposes a dynamic scheduling method based on continual transfer learning. Unlike traditional schedulers (e.g., Kubernetes) that require frequent offline retraining, our approach enables lightweight, online model evolution using incoming task streams—eliminating the need for periodic retraining and significantly reducing operational overhead. Empirical evaluation on the Google Cluster Trace demonstrates that the method maintains over 99% scheduling prediction accuracy while reducing average scheduling latency for affinity-constrained tasks and lowering overall system computational cost. The core contribution is the integration of a continual transfer learning mechanism into the scheduling decision loop, establishing the first affinity-aware, low-overhead, and high-temporal-fidelity adaptive scheduling framework. This advances scalability and dynamic responsiveness in large-scale HPC systems.

0 citationsRead paper