Fantasy football meets machine learning: the dynamic game case and a note on strategy
Deals with the development of threads synchronizing strategies based on the creation of concurrent «flat-combining» data structures as well as research of their performance. The paper considers «flat-combining» approach and its implementation in the library libcds, the development of thread synchronization strategy and its possible implementations. The efficiency of synchronization strategies usage is researched on the example of the open source library libcds. The research revealed the strategy with the lowest operation execution time on a container and the lowest amount of CPU resources, and identifies use cases of the developed strategies. A mechanism with the developed synchronization strategy to build concurrent data structures was implemented. The implemented strategies were integrated in the cross-platform open source library libcds.
To satisfy the stringent capacity and scalability requirements in the fifth generation (5G) mobile networks, both wireless access and backhaul links are envisioned to exploit millimeter wave (mmWave) spectrum. Here, similar to the design of access links, mmWave backhaul connections must also address many challenges such as multipath propagation and dynamic link blockage, which calls for advanced solutions to improve their reliability. To address these challenges, 3GPP New Radio technology is considering a flexible and reconfigurable backhaul architecture, which includes dynamic link rerouting to alternative paths. In this paper, we investigate the use of aerial relay nodes carried by e.g., unmanned aerial vehicles (UAVs) to allow for such dynamic routing, while mitigating the impact of occlusions on the terrestrial links. This novel concept requires an understanding of mmWave backhaul dynamics that accounts for: 1) realistic 3-D multipath mmWave propagation; 2) dynamic blockage of mmWave backhaul links; and 3) heterogeneous mobility of blockers and UAV-based assisting relays. We contribute the required mathematical framework that captures these phenomena to analyze the mmWave backhaul operation in characteristic urban environments. We also utilize this framework for a new assessment of mmWave backhaul performance by studying its spatial and temporal characteristics. We finally quantify the benefits of utilizing UAV assistance for more reliable mmWave backhaul. The numerical results are confirmed with 3GPP-calibrated simulations, while the framework itself can aid in the design of robust UAV-assisted backhaul infrastructures in future 5G mmWave cellular.
This research is based on Kaggle1 competition for Caterpillar Inc. Caterpillar Inc. sells a variety of construction and mining equipment. Each machine relies on a complex set of tubes. Tubes can vary across a number of dimensions, material, bends, length and other parameters. Caterpillar Inc. relies on a variety of suppliers to manufacture these tube assemblies, each having their own unique pricing model. The challenge is to predict supplier's tube price based on tube parameters. Caterpillar Inc. reviles novel field to apply machine learning technique. In this paper I reveal a good approach to solve this task. This solution ranked in a top 10% among more than 1300 contestants. Ranking was based on RMSLE and the solution achieved 0.218223. I found useful to ensemble various random forests  predictions with xgboost  library.
The main goal of this work is to present the developed research tool to find, investigate and analyze hidden dependences between parameters of the hardware/software platforms (such as influence of NUMA architecture, memory page size, etc) and the performance of block data processing algorithms. The new toolset (STAND) allows performance estimation and comparison of block data processing algorithms (for example, encryption/compression algorithms) running in kernel space. The primary application area of the developed technology and toolset is performance estimation and comparison of 'black box' libraries on particular hardware/software platform rather than research of mathematical or software implementation of algorithms. The main advantage of the presented toolset is that no source codes of algorithm implementation are needed (providing that an abstraction layer with known API is available). Linux operating system and computing nodes with ccNUMA architecture was selected as basic software/hardware platform. In this paper, the architecture of STAND is described. The methods for generating system load and comparison results for encryption algorithms AES (CBC), AES (CTR), and compression algorithms LZO, quicklz and bCodec are also presented.