ACCS-Canberra Node: Canberra Node of the ARC Centre on Complex systems

The main ACCS website click here

ACCS Scholarhips and ARC Linkage Scholarhips

Background

The ARC Centre for Complex Systems (ACCS) was established in 2003 with a total budget of $6 million dollars over 5 years and funded by the Australian Research Council (ARC) to undertake interdisciplinary research in the emerging discipline of complex systems science and engineering. ACCS conducts world-class basic and applied research on questions fundamental to understanding and managing complex systems. ACCS has leading Australian and overseas researchers in the area of complex systems. It is based at the University of Queensland (Brisbane) with nodes at Griffith University (Brisbane), Monash University (Melbourne), and the University of New South Wales at the Australian Defence Force Academy (Canberra) with associate investigators from other Australian and overseas research organizations. ACCS's partner organizations include Boeing, CSIRO, Hewlett Packard and Sun Microsystems. International collaborating organisations include France's Centre National de la Research Scientifique and the Indian Institute of Technology.

ACCS Canberra Node Chief Investigator and Contact Point

ACCS Canberra Node: Associated UNSW@ADFA Researchers:

ACCS Canberra Node: Associated International Researchers:

ACCS Canberra Node: ACCS-Funded Research Students:

ACCS Canberra Node: Associated Research Students:

ACCS Canberra Node: Projects

Topic: Evolution and Learning

Student Name: Mr. Lam Bui

An important paradigm in complexity is natural computation, which interprets natural phenomena as forms of computation. It also imitates nature (e.g. evolution) to derive computational methods for solving complex problems. Biological systems have been a rich source of ideas for solving complex computation problems (e.g. genetic algorithms, neural networks). Further, computational models have provided deep insights into biological processes (e.g. ALife).

Previous studies have shown a great deal of interaction occurring between the evolutionary level and the learning level (e.g. Baldwin effect). The objective of this project is to scrutinize this interaction in non-stationary evolutionary landscapes under complex genotype-phenotype mappings using non-trivial genetic encoding methods inspired by models from developmental biology such as: gradients, reaction-diffusion, chemical waves and genetic regulatory networks.

Topic: Multi-Agent Systems for Free Flight Air-Traffic Control

Student Name: Scholarship available

Air traffic control is one of the major bottlenecks preventing increased use of airspace and reduction in travel times. Free flight involves a fundamental shift from centralised control mechanisms (such as en-route air-traffic control) to localised control (whereby pilots take over primary responsibility for maintaining separation between aircraft). Major issues arise with respect to assuring safety and providing aviation services.

The objective of this project is to develop fully de-centralized control mechanisms to achieve safe free flight air-traffic control management system. A multi-agent approach will be used in conjunction with the Vortex software, a physics-based tool which can support building realistic virtual environments.

ALAR Related Projects

Modelling a Scenario Planning Operational Network (ARC Linkage, 0.5million ARC, 0.6 million Industry - NCR and IMIA)

Student Name: Scholarship available

This specific project on "modeling a scenario planning operational network" will address the need to model an operational business network to support scenario planning. This project will involve the development of a detailed agent model and protocols for the automated replication of the agent schemas, modelled to relevant stakeholders external to the business network. Modelling the operational network provides a unique set of challenges. Here the agents are more heterogeneous, including clients, suppliers, financial and other services, government agencies and so forth. The APAI will make use of the XML methods from another research stream in the overall project, but will take an important step forward in autonomous agent creation. The agents will be tested in the overarching framework developed in the integrating phase of the overall project.

Topic: Multi-Agent Systems for Planning: Modelling the Future as a Complex System

Student Name: Mr. Ang Yang (UNSW@ADFA)

Projecting a strategy into the future to predict the corresponding outcome has been an active area of research in machine learning and management. An equally important and more complex problem is to set a future goal and search for a set of conditions to achieve this goal. Modelling this problem mathematically is almost impossible when a large number of subjective factors influence the decision. In this thesis, we are planning to tackle this problem using a multi-agent complex system approach.

Currently, research in multi-agent systems (MAS) spans many areas of computer science, such as artificial intelligence, distributed systems, robotics and artificial life. An agent is any entity that can perceive its environment through sensors and act upon that environment through its actions. In many situations agents coexist and interact with other agents in several different ways. Examples include software agents on the Internet, robots playing soccer. war game simulation, and many more. Such systems that consist of a group of agents that can potentially interact with each other within a certain environment is called multi-agent systems. In this thesis, we will propose a flexible architecture of multi-agent systems to investigate the structure and evolution mechanisms of these systems. The architecture will be applied to different areas.

Topic: The Evolution of Ensemble of Artificial Neural Networks

Student Name: Miss Minh Ha Nguyen (UNSW@ADFA)

Most of the work in the Artificial Neural Networks (ANNs) literature concentrates on finding a single network. However, a single network found using the training set alone may not be the best network on the test set (i.e. it may not generalize well). The network can be either over-fitting the data or undertrained. Recently it has been found that by combining several neural networks, the generalization of the whole system could be enhanced over the separate generalization ability of the individuals. The main supportive argument for the performance enhancement of the ANN ensemble is that, since members of the ensemble possess different bias/variance trade-offs, a suitable combination of these biases/variances could result in an improvement in the generalization ability of the whole ensemble.

In the ANN ensemble literature, researchers have attempted to construct a population of individual neural networks and select the suitable ones to form the ensemble. A special branch of this literature body, we will name it the evolutionary ANN ensemble, applies evolutionary computation (EC) to evolve the population of neural networks. In the evolutionary ANN ensemble literature, no study has investigated the advantages of indirect encoding of the ensemble as a whole. By independently evolving and training the individuals, it is hard and so far unsuccessful to cooperate diversity into the system. It is pointed out by a number of researchers that the ensembles often did not express enough diversity as the methods claim to do. By evolving the whole ensemble as a whole, one has more control on the injection of the required diversity to the system.

Topic: Incremental Clustering

Student Name: Mr Damien Pumphrey (UNSW@ADFA)

Incremental clustering of data sets is becoming increasingly important as databases are beginning to process data online therefore patterns need to be extracted efficiently and accurately. By incorporating a number of machine learning techniques, exemplar and incremental learning, to develop a new algorithm that cluster data sets incrementally and accurately, it is possible to detect clusters of varying shapes, sizes and densities, process data incrementally thus reducing classification time, handle large data sets and computers with limited amounts of memory. Current clustering algorithms are unable to meet all these requirements thus limiting the types of applications they can be used in. I propose to develop a new incremental data mining algorithm that is able to handle online data processing and other applications that require the extraction of information incrementally by incorporating the above techniques.