ANU College of Engineering and Computer Science
School of Computer Science
|
ANU College of Engineering and Computer Science
School of Computer Science
|
|
|
||
|
Individual Project Description by Scott Sanner (2008)Title: Natural Language Query Answering
DescriptionEver wanted to ask your computer a question in natural language and have it respond with the answer you wanted? It sounds like a daunting task, but many of the foundations of such a technology are well understood. It is simply the process of inferring all possible relevant answers from an Internet full of content that is intractable. In this project, we view this inference task as one of reinforcement learning (i.e. how to learn to make a relevant sequence of inferences). In the end, our goal is to build a small prototype of this technology for natural language query answering in Wikipedia.Title: Finding Latent Variables in Reinforcement Learning
DescriptionReinforcement learning is the task of how to learn to make optimal sequential decisions when you can only sample experience from acting in a (simulated) environment. One of the greatest obstacles to practical reinforcement learning is finding the right features to use. In this project, we will examine automated feature induction methods via latent variable learning. The resulting reinforcement learning agent can be applied to one of the following tasks:
Title: Combining Machine Learning and Logic
DescriptionThere is great potential to exploit the powerful abstraction capabilities of logic for use in machine learning tasks. In this project, you would choose a task (natural language information extraction, query answering, probabilistic database inference, etc...) and implement a small learning and inference engine based on an appropriate logic (first-order, higher-order, description, or modal logic, etc...) and the recently studied structured learning tool of conditional Markov random fields (CRFs). A large library of Java tools will provide most of the underlying code needed to implement this project. Following implementation, the student will compare their work to state-of-the-art machine learning tools. This project offers the student the chance to learn about logic and machine learning, and how they can be combined for practical applications.Title: Mars Rovers and Traffic Controllers
DescriptionMarkov decision processes (MDPs) are a theoretical tool for modelling sequential decision making problems and their optimal solution. Recent advances in the theory of MDPs permit efficient solutions to problems with both continuous state and action spaces. Such models are highly appropriate for planning in both Mars Rovers and Traffic Controllers (just to name two examples). In this project, you would choose one of these problem domains (or perhaps another you can suggest) and implement an (approximately) optimal planning system for this task. This project offers the chance for the student to learn about the theory of optimal sequential decision making and its application to practical problems.Title: Scheduling for a High-performance Computing Cluster
DescriptionWe have a partner at the ANU who is looking for better process schedulers for their supercomputing systems. This requires performing some data analysis on their server logs to develop a model of process performance (based on a number of observed features) and then feeding this data to a constrained scheduler (which we have in-house). This project has great potential to make a substantial impact in the world of high-performance computing.
| |||||||||||||||||||||||||||||||||
