CS Faculty Candidate Presentation on Friday 4/4 (1-2PM, EB 1010)

[Dr. Stephen Blythe]

As many of you may have heard, the CS department is in the process of hiring a new faculty member to start in Fall of 2008. Each of the candidated will be doing an on-campus interview; the first is Dr. William Siever, who will be here this Friday (4/4). As part of his visit, he will be giving a presentation entitled "Reinforcement Learning, High Performance Computation, and Where They Meet"  from 1-2PM in EB1010 (abstract is listed below). After the meeting, he would like to meet with current students, so here is your chance to meet him and give your feedback on him. Remember - should we hire Dr. Siever, you could end up taking a class with him, so it would behoove you to give us your input!

  -Dr. Blythe


The Department of Computer Science
presents
an Invited Talk by Faculty Candidate

DR. WILLIAM SIEVER

Reinforcement Learning, High Performance Computation,
and Where They Meet

Abstract

Two areas of significant research in computer science are Artificial Intelligence and High Performance Computation. The speaker will introduce Reinforcement Learning (RL), an interesting niche of Artificial Intelligence, then detail some of the difficulties involved in the development of a typical high-performance algorithm. Finally, the speaker will discuss  how RL may be able to solve some of the issues of high-performance computation.

Reinforcement Learning (RL) is one of many artificial intelligence problem domains. RL is loosely based on the psychology principle of conditioning: the algorithm should learn behaviors that lead to good outcomes and avoid those that lead to bad outcomes. RL techniques are particularly powerful because they can be applied to problems with no known optimal solution or where the optimal solution changes over time. For example, RL algorithms have been shown to be very good at playing backgammon, robot soccer, and optimal system control.

Application of High Performance Computation (HPC) such as high-speed simulation requires careful software engineering design. For example, the development of a high-speed power system simulation requires in-depth understanding of A) power systems, B) numeric methods, and C) the system on which the simulation will be running. The choices made during development of such a simulation are often specific to both the problem and the architecture on which it will run, rendering the design non-portable.

The difficulty of tuning an HPC system to the compute architecture continues as modern computers become increasingly complex. For example, computers often have multiple cores, each of which is capable of some level of instruction level parallelism, as well as special purpose facilities such as graphics processing units (GPUs) that offer facilities for fine-grained, single-instruction multiple data parallelism.

An interesting area of future research is the development of "autotuners" - algorithms that use Reinforcement Learning to automatically adjust a simulation system to make the most effective use of the available computational resources.

Friday, April 4, 2008
1:00 -2:00 PM
EB 1010

[Shaun Martin]

For those of you who attended, what were your impressions of Dr. Siever?

[blacklee]

I really like him. What about you Shaun?