Research
RL4CO
An extensive Reinforcement Learning (RL) for Combinatorial Optimization (CO) benchmark. Our goal is to provide a unified framework for RL-based CO algorithms, and to facilitate reproducible research in this field, decoupling the science from the engineering. RL4CO is built upon: TorchRL: official PyTorch framework for RL algorithms and vectorized environments on GPUs; TensorDict: a library to easily handle heterogeneous data such as states, actions and rewards; PyTorch Lightning: a lightweight PyTorch wrapper for high-performance AI research; Hydra: a framework for elegantly configuring complex applications.
[Paper] [GitHub]
GraphSplineNets
We present GraphSplineNets, a novel deep-learning method to speed up the simulation of physical systems by reducing the grid size and number of iteration steps of deep surrogate models. Our method uses two differentiable orthogonal spline collocation methods to efficiently generate continuous solutions at any location in time and space. Additionally, we introduce an adaptive collocation strategy in space to prioritize sampling from the most important regions. GraphSplineNets improve the accuracy-speedup tradeoff in forecasting various dynamical systems with increasing complexity, including the heat equation, damped wave propagation, incompressible Navier-Stokes equations, and real-world ocean currents in both regular and irregular domains.
[Paper] [GitHub]
EvolveHyperGraph
In this paper, we propose a group-aware relational reasoning approach with explicit inference of the underlying dynamically evolving relational structures, and we demonstrate its effectiveness for multi-agent trajectory prediction. In addition to the edges between a pair of nodes (i.e., agents), we propose to infer hyperedges that adaptively connect multiple nodes to enable group-aware relational reasoning in an unsupervised manner without fixing the number of hyperedges. The proposed approach infers the dynamically evolving relation graphs and hypergraphs over time to capture the evolution of relations, The proposed approach infers the dynamically evolving relation graphs and hypergraphs over time to capture the evolution of relations, which the trajectory predictor uses to obtain future states.
[Paper] [GitHub]
Water Treatment Data Analysis
This study investigated the use of machine learning models, including traditional and deep learning approaches, for predicting both coagulant dosage and settled water turbidity in the water treatment process using six years of operating data. The study found that deep learning models, which process temporal sequential data, significantly improved prediction accuracies in response to changing dynamics of water treatment processes. The results emphasize the importance of collecting large datasets for modeling water treatment processes to capture rapid changes in raw water quality, thereby increasing prediction accuracies. The modeling results provide suggestions for model selection, data collection, and monitoring implementation in water treatment plants, which can enhance the accuracy of predictions and ensure high-quality treated water.
[Paper] [GitHub]Time-series Data Prediction
A graph attention multivariate time series forecasting (GAMTF) model was developed to determine coagulant dosage and was compared with conventional machine learning and deep learning models. The GAMTF model (R2 = 0.94, RMSE = 3.55) outperformed the other models (R2 = 0.63 - 0.89, RMSE = 4.80 - 38.98), and successfully predicted both coagulant dosage and settled water turbidity simultaneously. The GAMTF model improved the prediction accuracy by considering the hidden interrelationships between features and the past states of features. The results demonstrate the first successful application of multivariate time series deep learning model, especially, a state-of-the-art graph attention-based model, using long-term data for decision-support systems in water treatment processes.
[Paper] [GitHub]BSincNet
In this work, using the close relationship between time-series and its spectrum on frequency-domain, we propose a domain generalization method for time-series classification. Specifically, we quantify latent frequency spaces that captures the distinct characteristics of time-series data per each label and domains. Then, we use the filtered time-series on the assigned latent frequency spaces as augmented training sets for the model to make the model invariant classification up to each domain of the time-series. We demonstrate the proposed method improves the domain generalization performance on various time-series classification tasks.
[Paper]