A Scalable Machine Learning Approach to Go

Wu, Lin; Baldi, Pierre

doi:10.7551/mitpress/7503.003.0195

Cited by 26 publications

(49 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…However, these strategies sacrifice the expressivity needed to capture all-pair interactions among arbitrary nodes. Another recent work [46] proposes kernelized message passing that approximates the all-pair attention with linear complexity. However, this scheme introduces random feature maps that can lead to training instability.…”

Section: Preliminary and Related Workmentioning

confidence: 99%

“…Incorporation of Structural Information. For accommodating the prior information of the input graph G, existing models tend to use positional encodings [32], edge regularization loss [46] or augmenting the Transformer layers with GNNs [47]. Here we resort to a simple-yet-effective scheme that combines Z with the propagated embeddings by GNNs at the output layer:…”

Section: Simplifying and Empowering Transformers On Large Graphsmentioning

confidence: 99%

“…Scaling to Larger Graphs. For larger graphs that even GCN cannot be trained on using full-batch processing on a single GPU, we can use the random mini-batch partitioning method introduced by [46]. This incurs only negligible additional costs during training and allows the model to scale to arbitrarily large graphs.…”

Section: Simplifying and Empowering Transformers On Large Graphsmentioning

confidence: 99%

“…graph sizes, but sacrifices the expressivity for learning all-pair interactions. In contrast, NodeFormer [46] and SGFormer maintain attention computation over all N nodes in each layer while still achieving O(N ) complexity. However, unlike NodeFormer which relies on random features for approximation, SGFormer does not require any approximation or stochastic components and is more stable during training.…”

Section: Comparison With Existing Modelsmentioning

confidence: 99%

“…For further larger graphs, SGFormer is compatible with mini-batch training using large batch sizes, which allows the model to capture informative global information while having a negligible impact on performance. Notably, due to the linear complexity and simple architecture, SGFormer can scale to the web-scale graph ogbn-papers100M (with 0.1B nodes) when trained on a single GPU, two orders-of-magnitude larger than the largest demonstration of NodeFormer [46].…”

Section: Comparison With Existing Modelsmentioning

confidence: 99%

See 4 more Smart Citations

Compressor Rig Test Facility in Shanghai Jiao Tong University

Wu¹,

Yang²,

Sun³

et al. 2019

Proceedings of Global Power &Amp; Propulsion Society

View full text Add to dashboard Cite

From 2015, Turbofan Technology and Engineering Institute of Shanghai Jiao Tong University has made great efforts to set up a compressor rig test facility. Since this time, the facility has been used extensively. All systems for control, surveillance and data acquisition have been developed in-house, and it has been possible to make continuous improvements over the years. Also, the mechanical performance has been perfected. After several compressor tests of operation, the facility is now running at its optimum with a high degree of availability. Main characteristics of the facility are as follow. The maximum drive power is 4.3 MW, and the maximum mass flow is 50 kg/s. The inlet pressure can be controlled by inlet valve. The exhaust pressure is controlled by pneumatic butterfly valve, and 600 kPa absolute pressure maximum can be reached. The maximum exhaust temperature can reach 260 °C. Meantime, effective noise reduction to the control room and to the surrounding environment has been done. The inlet system and test object are set up in aerodynamic anechoic chamber, which is 20 m long, 18 m wide and 11 m tall. For the motor and gear box, they are insulated in a power room, and the wall of the power room is paved with sound attenuation material. The main part of the test facility is the control system and data acquisition system. A state-of-the-art control system with several safety features and data logging for later retrieval is set up. The stat-of-the-art data acquisition and evaluation system are based on NI data acquisition hardware. In the data acquisition and evaluation system, a stall alarm model and blade flutter detection model are integrated in the system. Now, this facility has been used for validating the booster design of ACAE. In this paper, the design of the facility, the control system and data acquisition and some test results are introduced in detailed.

show abstract

Section: Preliminary and Related Workmentioning

confidence: 99%

Section: Simplifying and Empowering Transformers On Large Graphsmentioning

confidence: 99%

Section: Simplifying and Empowering Transformers On Large Graphsmentioning

confidence: 99%

Section: Comparison With Existing Modelsmentioning

confidence: 99%

Section: Comparison With Existing Modelsmentioning

confidence: 99%

See 3 more Smart Citations

Compressor Rig Test Facility in Shanghai Jiao Tong University

Wu¹,

Yang²,

Sun³

et al. 2019

Proceedings of Global Power &Amp; Propulsion Society

View full text Add to dashboard Cite

show abstract

Design and Evaluation of Reinforcement Learning Based AI Agent: A Case Study in Gaming

Jayashree

Ramakrishnan

2017

Advances in Intelligent Systems and Computing

View full text Add to dashboard Cite

Mimicking Go Experts with Convolutional Neural Networks

Sutskever

Nair

Artificial Neural Networks - ICANN 2008

View full text Add to dashboard Cite

Abstract. Building a strong computer Go player is a longstanding open problem. In this paper we consider the related problem of predicting the moves made by Go experts in professional games. The ability to predict experts' moves is useful, because it can, in principle, be used to narrow the search done by a computer Go player. We applied an ensemble of convolutional neural networks to this problem. Our main result is that the ensemble learns to predict 36.9% of the moves made in test expert Go games, improving upon the state of the art, and that the best single convolutional neural network of the ensemble achieves 34% accuracy. This network has less than 10 4 parameters.

show abstract

A Scalable Machine Learning Approach to Go

Cited by 26 publications

References 7 publications

Compressor Rig Test Facility in Shanghai Jiao Tong University

Compressor Rig Test Facility in Shanghai Jiao Tong University

Design and Evaluation of Reinforcement Learning Based AI Agent: A Case Study in Gaming

Mimicking Go Experts with Convolutional Neural Networks

Contact Info

Product

Resources

About