Projects

We have experience with a wide variety of Bayesian network applications. Past and current projects include:

• Modelling Threats to Species

  Victoria's Department of Sustainability and Environment (DSE) uses the Actions for Biodiversity Conservation system for recording information on threatened species across the state. We created a software component (known as ABC-BN) that fully integrates into this database, allowing species managers to explicitly model the impacts of threats on species and the effectiveness of management actions used to mitigate those threats.

  ABC-BN provides a comprehensive set of tools for creating models of the factors affecting a species and exploring the predictions of those models. To keep things simple for managers, a template model structure is used that captures the fundamental elements important to every species' situation: namely, management actions, threats and population factors (such as species numbers or range). For any given species, managers are free to include in their model whatever actions, threats and population factors they believe to be important. Managers can then add links to indicate which actions affect which threats, and which threats impact upon which population factors. Parameterisation of the network is performed using natural language questions and prompts, with the ability to switch to tables and numbers for those more comfortable with BN concepts. After the models are reviewed and accepted, stakeholders can experiment with the created models by entering either certain or uncertain evidence (in the form of likelihoods) and produce reports indicating which actions and other factors are most likely to improve the status of threats and, ultimately, the status of the species.

• Aggregating Expert Forecasts

  The US government organisation IARPA established the ACE (Aggregative Contingent Estimation) program in order to develop and improve methods for eliciting conditional predictions from experts and also for aggregating those predictions into a single prediction. One of the teams participating in that program is a group called DAGGRE, based at George Mason University. Together with ACERA, we are assisting DAGGRE with the development of their combinatorial prediction market software, now based on an underlying BN back-end, along with their conditional and unconditional elicitation processes. We additionally created information gathering and elicitation software for ACERA's own elicitation and aggregated prediction experiments.

• Victorian Western Grasslands Reserves

  The Victorian Government is reserving 15,000 hectares of land to the west and north-west of Werribee as a way of protecting endangered grasslands and offsetting the ecological costs of urban growth to the north and west of Melbourne. The DSE wishes to provide Parks Victoria with an adaptive management process for these grasslands, which will involve modelling, predicting, monitoring and revising the model in repsonse to discrepancies that arise in the review. DSE has chosen Bayesian networks as the modelling technology, and we are building the initial models to be used in this process, as well as providing training.

• Power Pole Maintenance

  Western Power manages over 700,000 power poles in Perth and surrounding areas, many of which are nearing the end of their service lives. Western Power has recently embarked on a project aimed at improving the safety of these poles, while minimising the maintenance cost. The company settled on Bayesian networks as the best approach to model the health of the wood poles, given their strong risk management capabilities. We are providing Western Power with modelling advice and performance-focused BN programming solutions to assist in managing the risk of power pole failure.

• Anomaly Detection

  We are creating software for the Defence Science and Technology Organisation based on our machine learning software CaMML and earlier research work to detect anomalous behaviour in track data sets of various kinds (such as ships and cars). This project involves mining large time series data sets in order to produce causal models that yield an understanding of normal behaviour, against which anomalies can be assessed. The emphasis in the current project will be to allow users of the software to create and test models and to provide detection in an online environment.

• Biosecurity

  We are developing models for New Zealand's Scion research institute to minimise the risk of pests appearing in wood chips due for export. The aim is to sufficiently reduce the risk of attracting pests throughout the cutting, transportation and storage stages, such that costly and environmentally-unfriendly fumigation can be avoided prior to export.

Projects the principals have completed successfully (prior to the establishment of Bayesian Intelligence in 2008) include:

• Ecological Modelling
  

  We have done a number of ecological modelling projects, including

  • providing intelligent decision support for the assessment of ecological risk associated with irrigation systems in the Goulburn-Broken catchment;
  • providing intelligent decision support for tropical seagrass in the Great Barrier Reef;
  • assessing water quality in Sydney Harbour.

• Bureau of Meteorology
  

  We've done two projects with the Bureau. A prototype project in 2000 used the substantial amount of data gathered in the lead-up to the Sydney Olympics to model seabreezes. The study used a combination of expert elicitation and CaMML (Causal Discovery via MML) to find Bayesian networks that improved upon existing BOM prediction systems. A second project has developed an ontology for weather forecasting and a tool for examining and manipulating the ontology; the tool provides a form of object-oriented Bayesian network design, by allowing default subnetworks to be associated with weather objects, which can be extracted, composed and refined. This second project also involves the development of Bayesian networks for decision support in forecasting fog and hailstorms.

• Intelligent Tutoring Systems
  

  We have used Bayesian networks to develop an intelligent tutoring system, DecSys, for children learning decimal arithmetic, using simple computer game environments. The BN models the children's understanding of arithmetic and especially their misconceptions, allowing sensible automated judgements of what material to present or re-present next. (See DecSys for more information.)

  This project was undertaken in collaboration with researchers at the University of Melbourne.

• Bayesian Poker
  

  Our Bayesian poker player (BPP) uses Bayesian networks to model opponents and to model hands in Texas Hold'em Poker. BPP was entered into the inaugural and subsequent world automated poker playing competitions at the American Artificial Intelligence Conference (AAAI), from 2006. A simple GUI interface allows people to play against BPP via the internet.

• Cardiovascular Risk Assessment
  

  Using both expert knowledge and longitudinal data we developed a Bayesian network model of coronary heart disease, TakeHeart II, which supports risk assessment for individuals with and without treatment interventions. As a part of this project we developed a scripting language to support development of GUI interfaces sitting above a Bayesian network which serves as the inference engine. This allowed us to place a superior human-computer interface on top of a model that performs as well as the best published models. This modularity will also allow a simple should better Bayesian network models become available.

• Biomedical Engineering
  

  We developed dynamic Bayesian networks for ambulation monitoring of the elderly and to diagnosis falls. The monitoring is performed using two kinds of sensors: foot-switches, which report steps, and a mercury sensor, which is triggered by a change in height. The networks issue an alert when a fall is diagnosed.

• Software Architecture Design
  

  We have used Bayesian networks for change impact analysis in software architecture design.

  Research into system design rationale in the past has focused on the representation of design deliberations and has omitted the connections between design rationales and design artefacts. Without such connections, designers and architects cannot easily assess how changing requirements or design decisions may affect the system. In this project, we introduced the Architecture Rationale and Element Linkage (AREL) model to represent the causal relationships between architecture elements and decisions. AREL was in turn modelled as a Bayesian Network, capturing the probabilistic dependency relationships between the architecture elements and decisions in such an architecture design model. We demonstrated that such probabilistic modelling enables architects to quantitatively analyse (i.e. predict and diagnose) the impact of change in either the requirements or the design.

  This project was undertaken in collaboration with researchers at Swinburne University.

References

References are available upon request.