Compact Vehicle Roaming
Philip Kilby & Charles Gretton
In the Vehicle Routing Problem, a set of customers is
visited by a fleet of vehicles so as to minimize delivery costs. Constraints
such as vehicle capacity, and visiting customers within a nominated time window
must be observed. This is a problem with huge practical application.
In real-life situations, fleet managers often express a
preference for routes that are "compact": the area served by a
vehicle should not overlap with areas served by other routes. Routes then
define a non-overlapping "service area" for each driver.
This project will look at methods for constructing routes
that have this feature. The work will involve ideas from artificial
intelligence, optimisation, and also from computational geometry.
Future Energy Systems
The world we live in is continuously changing and natural disasters are unfortunately increasing. The way we handle energy, its generation and consumption will affect the world we will leave behind for future generations.
The aim of this project is to develop efficient algorithms for a range of critical energy related problems for tomorrow's smart grid. The subject includes robust smart grid design and power restoration in disaster management.
The goals of this project are to develop reliable and scalable methods for solving complex energy related problems featuring nonlinear constraints, discrete decisions, and uncertainty.
- Real impact on real world problems.
- Expertise in Mathematical Optimisation.
- Direct contact with industry.
Generic Reinforcement Learning Agents (GRLA)
Marcus Hutter & Peter Sunehag
Agent applications are ubiquitous in commerce and industry, and the sophistication, complexity, and importance of these applications is increasing rapidly; they include speech recognition systems, vision systems, search engines, planetary explorers, auto-pilots, spam filters, and robots [RN03]. Existing agent technology can be improved by developing systems that can automatically acquire during deployment much of the knowledge that would otherwise be required to be built in by agent designers. This greatly reduces the effort required for agent construction, and results in agents that are more adaptive and operate successfully in a wide variety of environments.
Goals of this project: Technically, the project is about a recent general approach to learning that bridges the gap between theory and practice in reinforcement learning (RL). General-purpose, intelligent, learning agents cycle through sequences of observations, actions, and rewards that are complex, uncertain, unknown, and non-Markovian [RN03].
On the other hand, RL is well-developed for small finite state Markov decision processes (MDPs) [SB98]. Extracting the right state representations out of bare observations, that is, reducing the general agent setup to the MDP framework, is an art that involves significant effort by designers. The project is to investigate (by simulations or theoretical) recent models [Hut09] that automate the reduction process and thereby significantly expand the scope of many existing RL algorithms and the agents that employ them.
Human Knowledge Compression Contest(HKCC)
Being able to compress well is closely related to intelligence as explained below. While intelligence is a slippery concept,file sizes are hard numbers. Wikipedia is an extensive snapshot of Human Knowledge.
If you can compress the first 100MB of Wikipedia better than your predecessors, your (de)compressor likely has to be smart(er). The intention of the Human Knowledge Compression Prize [Hut06] is to encourage development of intelligent compressors/programs.
Goals of this project. Some of the following four subgoals shall be addressed:
- Get acquainted with the current state of the art compressor and in particular with the prize winning paq8hpX series, and write a comparative survey.
- Develop and test novel compression ideas.
- Integrate them into one of the state-of-the-art compressors.
- Investigate alternative performance measures that take the compressor more seriously into account (rather than only the decompressor).
Latent Support Vector Machine for Distributed Solar Prediction
The Distributed Solar Prediction project at ANU and NICTA has collected a large volume of data from rooftop photovoltaic (PV) panels distributed across Canberra. With the power output recorded at every 10 minutes, machine learning technologies can be applied to predict the PV output in the future 30 minutes to 2 hours, which is useful for industry in many ways. The key feature of the learning paradigm is to fuse the data from multiple sites at multiple times, in order to improve the prediction accuracy at each site.
The goal of the summer project is to address an important practical issue: frequently data loggers break down and we are left with missing data. This problem is exacerbated when our model is based on multiple sites. A natural solution in machine learning is to use latent support vector machines, which learn the forecaster concurrently with the inference of the missing data. It can also be easily scaled up to very large datasets.
Machine Learning in the Cloud
The Protocols and Structures for Inference (PSI) project (see http://psikit.net ) is consists of a specification and prototype implementation for presenting machine learning data, learners, and predictors as RESTful web services. The aim of the specification is to make it easier for people to make use of machine learning tools to analyse their data and the allow interoperability between machine learning components over the web.
The aim of this project is to build more PSI-compatible services by writing adaptors for existing algorithms in machine learning toolkits such as scikit-learn, Shogun, and Orange or by wrapping existing ML services such as Google Predict, BigML, and wise.io. The newly developed services will be deployed on the Amazon Web Services platform (using funds from an Amazon in Education grant). A student completing this project will gain experience in both web service development and applied machine learning.
Machine understanding of images
Machine understanding of images is one of the long term goals of AI. In this project you will work on advanced machine learning algorithms for scene understanding.
Given a set of training images annotated with information about what objects and background regions are present, the algorithms learn to recognise these objects and background regions in new
Students should have a background in machine learning and strong C/C++ or Matlab programming skills.
Mathematical Foundations of Artificial Intelligence (MFAI)
Marcus Hutter & Peter Sunehag
The first decade of this century has seen the nascency of the first mathematical theory of general artificial intelligence. This theory of Universal Artificial Intelligence (UAI) has made significant contributions to many theoretical, philosophical, and practical AI questions.
In a series of papers culminating in book [Hut05], an exciting sound and
complete mathematical model for a super intelligent agent (AIXI) has been
developed and rigorously analyzed. The model is actually quite elegant and can
be defined in a single line.
Goals of this project: The fundamentals of UAI are already laid out, but there are literally hundreds of fundamental theoretical/mathematical open questions [Hut05,Hut09] in this approach that
have not yet been answered.
On the Foundations of Inductive Reasoning (FIR)
Humans and many other intelligent systems (have to) learn from experience, build models of the environment from the acquired knowledge, and use these models for prediction. In philosophy this is called inductive inference, in statistics it is called estimation and prediction, and in computer science it is addressed by machine learning.
The problem of how we (should) do inductive inference is of utmost importance in science and beyond. There are many apparently open problems regarding induction, the confirmation problem (Black raven paradox), the zero p(oste)rior problem, reparametrization invariance, the old-evidence
and updating problems, to mention just a few.
Solomonoff's theory of universal induction based on Occam's and Epicurus' principles, Bayesian probability theory, and Turing's universal machine [Hut05], presents a theoretical solution [Hut07].
Goals of this project.
- Elaborate on some of the solutions presented in[Hut07]
- Present them in a generally accessible form, illustrating them with lots of examples.
- Address other open induction problem.
Optimal path-finding in octile grids
Given a starting point and a target point on a map equipped with obstacle, the goal is to find the shortest path that links these two points. Eight moves are available, in straight (for instance, heading north or east) or diagonal (for instance south-east) direction.
This research project is concerned with optimal path-finding in octile grids.
A well-known solution to this problem is to use A* but, as it turns out, it is possible to develop specialised algorithms that perform much faster than that. We developed an algorithm called JPS  that eliminates (on-line) most of the symmetries in the map and outperform significantly A*. More recently, we developed an extension of JPS called JPS+  that uses preprocessing to improve the performance even further. JPS+ indeed does runs faster than JPS, but this comes at a price of flexibility since any change in the map requires to update the structure built at preprocess time.
JPS+'s preprocessing time is quite fast, but we would like to be able to update the structure without performing the full precomputation. The goal of this project is to identify what part of the precomputed structure needs to be updated, and to check the cost of repairing the structure.
Planning, problem solving and acting.
Planning problems come in many disguises, from control airport ground traffic, elevators, or high-performance printers, to computing genome similarity, discovering faults in protocols or data streams, or even generating a storyline.
The aim of planning research (as in many other branches of AI) is to construct domain-independent ("universal") solutions for this kind of problem. That is, rather than solving each application problem individually, a general AI planning system should be able to solve any one of them, provided a formal specification of the problem as input.
A list of some current projects is available at http://cecs.anu.edu.au.virtual.anu.edu.au/projects/500. Please contact me if you would like to know more about any of them.
Universal Artificial Intelligence (UAI)
Marcus Hutter & Peter Sunehag
The dream of creating artificial devices that reach or outperform human intelligence is an old one. Most AI research is bottom-up, extending existing ideas and algorithms beyond their limited domain of applicability. The information-theoretic top-down approach (UAI) pursued in [Hut05] justifies, formalizes, investigates, and approximates the core of intelligence: the ability to succeed in a wide range of environments [LH07]. All other properties are emergent.
Back to Top
Recently, effective approximations of UAI have been derived and experimentally investigated [VNHUS11]. This practical breakthrough has resulted in some impressive applications. For the first time, without providing any domain knowledge, the same agent is able to self-adapt to a diverse range of interactive environments. For instance, it is able to learn from scratch to play TicTacToe, Pacman, Kuhn Poker, and other games by trial and error.
These achievements give new hope that the grand goal of Artificial General Intelligence is not elusive.
Goals of this project: The theoretical [Hut05],philosophical [LH07], and experimental [VNHUS11] foundations of UAI are already laid out, but plenty remains to be done to solve the AI problem in practice. The complexity of the open problems ranges from suitable-or-short-projects to full
PhD theses and beyond.