Dynamic Neighborhood Construction for Structured Large Discrete Action Spaces

Authors: Fabian Akkerman, Julius Luy, Wouter van Heeswijk, Maximilian Schiffer

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the performance of our method by benchmarking it against three state-of-the-art approaches designed for large discrete action spaces across three distinct environments. Our results show that DNC matches or outperforms state-of-the-art approaches while being computationally more efficient.
Researcher Affiliation Academia 1 University of Twente (7500 AE Enschede, The Netherlands) 2 School of Management, Technical University of Munich (80333 Munich, Germany) 3 Munich Data Science Institute, Technical University of Munich (85748 Garching, Germany)
Pseudocode Yes Algorithm 1 Dynamic Neighborhood Construction
Open Source Code Yes The code for running DNC and all benchmarks on the studied domains can be found in the attached supplement and at: https://github.com/tum BAIS/ dynamic Neighborhood Construction.
Open Datasets Yes First, we consider a maze environment, which is used as a testbed in state-of-the-art works on LDAS (cf. Chandak et al., 2019a). Second, we consider a joint inventory replenishment problem (cf. Vanvuchelen et al., 2022). Third, we consider a well-known dynamic job-shop scheduling problem (see, e.g., Zhang et al., 2020a; Wu & Yan, 2023).
Dataset Splits No The paper describes training within reinforcement learning environments, using 'episodes' and 'timesteps', but does not specify explicit training/validation/test dataset splits with percentages or sample counts in the way typical for supervised learning on fixed datasets.
Hardware Specification Yes Our experiments are conducted on a high-performance cluster with 2.6Ghz CPUs with 56 threads and 64gb RAM per node.
Software Dependencies No The algorithms are coded in Python 3 and we use Py Torch to construct neural network architectures (Paszke et al., 2019). It mentions Python 3 but does not specify the version of PyTorch, which is a key software component.
Experiment Setup Yes In this section, we detail the hyperparameter settings used across the different environments. To this end, we provide an overview of hyperparameter settings in Table 2 and discuss specific settings that we use over all environments.