High-Performance Rendering of Interactive Decision Support Visualizations in
Network Restricted Environments, 10-R8296
Principal Investigator
John G. Whipple
Inclusive Dates: 04/02/12 – 08/02/12
Background — Decision support systems collect large amounts of historical data, but making decisions based on the data is difficult without summarizing visualizations. Traditional decision support visualizations allow the user to manipulate settings to view the data in a way that might make more sense; however, a round trip to a back-end server, or more recently "the cloud," is required before the visualization can be re-rendered. This is the accepted visualization paradigm; nevertheless, there are two intrinsic flaws with relying on a remote computer. First, network connectivity may not be available due to practical or procedural reasons. Second, including a network round trip limits the speed that the user can modify settings and see an updated visualization. This project investigated the high-speed rendering of interactive decision support visualizations that relied solely on the processing power of the computer hosting the visualization.
Approach — The first step to achieve the project objectives was to obtain a time-series predictive model that could be used as a basis for two custom visualizations. Data was randomly generated based on a known function that reflected a generic two-party negotiation data set. The generated data was then modeled using two different modeling techniques: neural network and Markov Chain Monte Carlo (MCMC). The focus of this research was not the quality of the model, but the quality of the visualizations and the speed at which they could be rendered. Next, the application algorithms for the modeling techniques as well as the visualizations were developed using HTML5 and JavaScript. A rendering test script was used to capture performance metrics. Test results were obtained from Internet Explorer®, Firefox® and Chrome®. An Apple® iPad was also included in the testing. An average frame rate of 15 FPS or greater was considered to be a success, as this frame rate is typical of mobile device refresh speeds and is unperceivable to the human eye.
Accomplishments — The neural network model did not pose a performance problem for any platform or browser. The frame rate consistently peaked at 60 FPS regardless of how the user interacted with the visualization. The calculations required of the MCMC were significantly more complex. The accuracy of the MCMC depended on the number of simulation iterations. It was discovered that all modern browsers running on modern hardware were able to maintain the 15FPS metric while iterating through the MCMC simulation 1,500 times. As a result of this research, it is now known that it is practical to use neural networks and, to an extent, MCMC using only Internet browser technology.
