Presentations, Tutorials and Archived Program

Wayfinding is the interplay between environmental factors such as landmarks, signage and pathways and human factors such as cognitive mapping abilities, language, cultural background and biological factors to allow the successful navigation of a site effectively and without confusion. An understanding of wayfinding, including the need for effective and efficient flow of people, are particularly important in the setting of transport terminals, particularly airports. In this setting, passengers can be nervous, time-constrained, come from different cultural backgrounds, are unfamiliar with the language used, are inexperienced in the travel process and are in an unfamiliar environment. A wayfinding system based on principles and research is in place in an airport, may negate some of the negative aspects experienced by passengers and in turn improve passenger satisfaction and experience. The need to understand wayfinding has been of interest to the fields of cognitive and computer science with these respective disciplines providing insight into the processes involved in wayfinding.

This paper proposes a Bayesian Network approach to wayfinding, which will bring together the human and environmental factors of wayfinding, and will incorporate previous research in the area. Additionally, the proposed Wayfinding BN model will be applied to a "Dashboard for Wayfinding". This dashboard will assist airport operators in determining the effectiveness of the current wayfinding environment in their terminals and allow for comparison to proposed changes in the terminal. The Wayfinding BN and Dashboard will be used to analyse the Common User Area at the Brisbane Domestic Airport before and after proposed changes in signage and wayfinding elements. more >

Airport terminals rely on the knowledge of experts to operate efficiently, but when an expert leaves they take their knowledge with them. A previous attempt to resolve this problem was to establish a set of metrics that could be used to assess the performance of the inbound passenger facilitation system independently. However, considering metrics independently is a sub-optimal method of predicting the behaviour of a complex system such as in an airport.

Airports represent a situation common in operational management in which both extensive expert knowledge and overwhelming data are available. Bayesian Network modelling allows the combination of expert knowledge and existing metrics to assist machine learning of model structures and parameters, and the use of trained and untrained models to learn more about the nature of existing expert knowledge and metrics. We demonstrate the dynamic object oriented Bayesian Network of the inbound passenger facilitation process that has been devised through expert elicitation. The application of rigorous validity tests initially strengthened the structure of the network as the model was adjusted to conform to their requirements. The model was then used to determine the importance of uncollected data to the validity of the model using a random number generation method, allowing us to both determine the comprehensiveness of the existing metric framework and to determine the value of gaining new measures.

The model presented will be used to combine the opinions of multiple experts with dynamically updated data for use in a range of applications for airport stakeholders. more >

Bayesian networks are generic modeling tools for analysing, understanding and predicting causal systems. When Bayesian nets are used for prediction, they effectively become classifiers, classifying some target variable whose values are inside or outside some class of interest. Naive Bayes models are popular Bayesian network classifiers, for example.

Discretization of continuous variables is an important technique for classification. The most common method is global entropy-based MDL discretization (Fayyad and Irani, 1993). We will critique common discretization methods, in particular criticizing unsupervised methods and methods (such as using predictive accuracy) which ignore calibration. We will introduce a number of metrics, such as cumulative expected utility and Bayesian information reward, which reward calibration as well as accuracy. Finally, we will describe a genetic algorithm which finds optimal discretizations using these metrics. more >

Whether a Bayesian Network (BN) is constructed through expert elicitation, from data, or a combination of both, evaluation of the resultant BN is a crucial part of the knowledge engineering process. One kind of evaluation is to analyze how sensitive the network is to changes in inputs, a form of sensitivity analysis commonly called ``sensitivity to findings'' (the term used in the Netica BN software). The properties of d-separation can be used to determine whether or not evidence (or findings) about one variable may influence belief in a query variable, given the BN structure only. Once the network is parameterised, it is also possible to measure this influence. Entropy is the common measure of how much uncertainty is represented in a probability mass, hence entropy reduction is one possible metric of change. A second measure of uncertainty sometimes used is the variance.

Given such a metric of change, when evaluating a BN, it is common to rank evidence nodes for either a maximum such effect or the average such effect. We show that this ranking tends to reflect the structural properties in the network: the longer the path from the evidence node to the query node, the less/lower the influence, while the influence increases with the number of such paths. This raises, the question: how useful is the ranking computed with the parameterised network, over and above what could be discerned from the structure alone? We propose a metric that ranks the influence of query nodes based on the structure of the network alone, and show how the interest from an evaluation perspective is how much the parameterisation changes the ranking. We illustrate the practical use of this on a range of networks. more >

Many algorithms have been developed for the purpose of learning static Bayesian networks but there has been little investigation into the learning of dynamic networks. The causal discovery program CaMML is able to take advantage of hard and soft prior expert knowledge in the learning process. Here we describe how these representations of prior knowledge can be used to turn CaMML into a promising tool for learning dynamic Bayesian networks. more >

Wildfires can result in large losses of property when they encounter the urban interface. Fire management agencies invest in a range of preventative and responsive management actions to reduce the risk of loss. Budgets for these actions are limited and therefore it is vital that these agencies invest in strategies that provide the greatest reduction in risk of loss.

Here we examine the interactive effects of urban planning, a preventative management approach (community education regarding house preparedness) and a responsive management action (active fire suppression) in a Dynamic Bayesian Network. Data for the model are derived from a combination of a process based model and a novel rapid expert elicitation method.

The elicitation process was successful with consistent trends between the groups involved. Current approaches to community education were not found to result in significant shifts in the behaviour of home owners. Trends in suppression effectiveness were also observed, with a strong dichotomy in participant's views of the role of aerial suppression. Gaps in the current understanding of processes were also identified.

The resultant DBN shows that, according to our experts investment in suppression resulted in the greatest reduction in house loss with the community education strategies tested having a relatively small effect. However, an idealised community education strategy which resulted in a greater level of house preparedness had a significant interaction with suppression actions. Urban planning practices were found to reduce the extent of house loss only when combined with investment in active fire suppression. more >

This paper aims at enhancing decision making in elite sports by means of probabilistic reasoning. Making informed decisions are vital for successful performance outcomes in the dynamic elite sports domain. Bayesian belief networks are excellent candidate machine learning methods to support such decision making since they can model the uncertainty involved in the sports competitions mainly due to the open environment in which athletes interact with teammates and/or opponents. BBNs can also combine different sources of evidence which is particularly useful in the sports domain where overall performances may depend on heterogeneous performance indicators (e.g., the stroke rate and start time in swimming). We have, therefore, utilized BBNs in a few sports to: i) understand the performance patterns of successful elite athletes, ii) predict future performances before and during competitions, and iii) find the limitations of current Bayesian learning methods as applied to the elite sports domain. As one instance, we have modeled the relationships between the ranking of cyclists in each component of the six-event track cycling omnium and their overall standing determined using the summation of the cyclist's ranking in all the six events. This assists coaches in finding whether sprint or endurance riders may have the greater chance of medal winning since each component of the omnium requires different capabilities. Despite our successful utilization of BBNs in a few sports so far, we have found two main limitations in the existing conditional probability learning methods when used in this domain: i) record (unseen) performance measures do not result in predicting a high likelihood of success which is counter-intuitive and ii) the importance of recent performance results as compared with older results are undermined and cannot be promoted even using the fading mechanism. These two limitations are the main avenues of our future research in this domain. more >

Western Power manages over 600,000 wood power poles across the south west of Western Australia. Pole failure can result in catastrophic consequences such as bushfire, electrocution and disconnection of an essential service. An effective wood pole management strategy requires that each pole is replaced at a time that maximises its useful life while minimising likelihood of failure and associated consequences of failure.

The useful life of a pole varies significantly due to factors such as wood type, internal and external rot, termite attack and soil conditions. The consequences of failure vary due to location factors such as connected customers, bushfire hazard and human activity.

A Bayesian belief network has been developed to assess the likelihood and consequence of failure of each pole. Likelihood of failure is based on the interaction of two probability distributions, one for strength and one for wind load. Each probability distribution is calculated from its associated influencing factors. Models use a combination of engineering equations, expert opinion and condition inspection data.

Mean consequence is calculated by Bayesian model to estimate a Normal Loss Expectancy and a Maximum Foreseeable Loss using data associated with the pole's location. A risk index, similar to an insurance premium, is calculated for each pole using its expected probability of failure in a given year and its mean expected cost of failure. more >

The Coorong, Lower Lakes and Murray Mouth Ramsar site is ranked as one of Australia's most important wetlands for migratory shorebirds. The site regularly supports over 100,000 waterbirds in summer, when large numbers of international migrants visit to forage on local prey resources. The distribution and abundance of waterbirds at this site is largely regulated by water flows from the River Murray and associated ecological conditions within wetland habitats. Between the early 2000s-2009, prolonged drought and upstream diversion of River Murray water resulted in a cascade of adverse ecological changes in the CLLMM. Water levels in the Lower Lakes fell below sea level, exposing harmful acid-sulfate soils, and salinity in the Coorong South Lagoon increased to >200ppt (modelled natural is 80ppt). These unprecedented conditions had a negative impact on the abundance and distribution of waterbirds as well as the fish, macroinvertebrate and plant species that make up much of their diet. In order to better understand the impact of the site's hydrology on the availability of waterbird habitat, we developed Bayesian models that enable managers to predict the consequences of ecological change for waterbirds. These species-specific models characterise cause and effect relationships between habitat components and a particular measure of waterbird habitat. We demonstrate the use of these habitat models as tools for the effective management and conservation of waterbirds. more >

As much of the natural vegetation in South Eastern Australia has been cleared since European settlement, a large portion of natural resource management in this geographical area focuses on attempting to restore biodiversity to degraded or simplified ecosystems. Despite large investments into this process, there is ongoing uncertainty about the effectiveness of many available management programs. To manage this uncertainty, natural resource managers are being encouraged to create models that are updated utilising an adaptive management approach. Rumpff et al. (2011) describe a site-scale model for use in the management of non-riparian grassy woodlands in the Goulburn Broken catchment. To capture the uncertainty involved in managing woodlands in different states of condition, the model was developed as a state-transition Bayesian network (ST-BN). We describe an architecture for a model that scales to a larger woodlands area by integrating the initial attribute and environment variables from a GIS, a weed dispersal dynamics model and an object-oriented Bayesian network (OOBN). In this paper we describe the process followed to convert the static ST-BN to an object-oriented dynamic Bayesian network (OODBN); the expert elicitation process used to develop the (ST-OODBN) structure building on the work described by Nicholson and Flores (2011); the process of prototyping a OODBN weed dispersal sub-model, and some of the challenges associated with acquiring appropriate GIS input layers. We present a scenario-based evaluation of the prototype ST-OODBN model and the weed dispersal sub-model. more >

One of the often reported benefits of Bayesian Networks (BN) is the simple graphical display offered by software packages and ease of exploring the model by users. My experience has found that this is not necessarily true with considerable detail hidden in the conditional probability tables of the models and users struggling to understand the reasons for model outcomes. While an uninformed user will receive immediate feedback from changing the state of a BN node, they will often not understand the reasons for the changes without additional feedback from an expert. This lack of narrative with respect to model outcomes therefore limits the use of the model as a standalone tool. My research aims to develop BNs with indigenous Australians to provide tools for disseminating indigenous ecological understanding. To achieve this I have been developing approaches to help users explore the understanding held in BNs using web based experiences linked to the BN. In this presentation I will present some of the lessons learned communicating the model understanding to a wide audience from scientists to the general public. more >

Increasingly, there is public, industry and government demand for measurement and quantification of impact and demonstration of sustainability practice, across the entire spectrum of the dairy industry: farm, factory and market. Although there has been significant progress towards measuring and tracking a range of individual indicators, there is a need to understand the complexity of sustainability through an integrated framework. The scorecard model offers dairy stakeholders at the farm, factory and consumer level a better understanding of the inter-relationships between key sustainability variables and the potential cause and effect impacts resulting from changes in key variables. We propose a sustainability scorecard model to measure the current Triple Bottom Line (TBL: economic, social and environmental) performance of the dairy industry in Australia using a Bayesian Network (BN) approach. The design of the proposed scorecard will collectively and individually measure the three domains of interest for dairy stakeholders: farm, factory and market, against TBL factors. The sustainability scorecard has key indicators for each of the three TBL, repeated for farm, factory and market. We illustrate the quantification of the BN models and include ratings from a dairy stakeholder meeting to determine the relative importance of the TBL within farm, factor and market, before summarising them in the final overall sustainability score. The proposed scorecard is a transparent and independent measure of change, illustrating the "ripple" effect of the "complex system" of indicators and interrelated impacts, identifying ways to increase overall sustainability while mitigating negative factors. more >

Movement of agricultural produce often creates risks of moving pests that are present in the origin but not the destination. In such cases, trade is conditional on application of pest risk management measures, which must be consistent with an international policy framework. Conventionally the measures are single, such as treatment with a pesticide of high efficacy such as methyl bromide. Such single measures can be problematic for reasons such as damage to the product, or high cost, or treatment failure in a few cases resulting in suspension of the whole trade. Alternatively, a "systems approach" (SA) can combine two or more independent treatments, which combined achieve the required level of risk mitigation. While SAs may offer several advantages, such as lower cost or higher robustness, it can be difficult or impossible, with a multifactorial system in a context of high variability, to collect the data required to formulate and negotiate a trade agreement. Thus, SA adoption has fallen short of its perceived potential. Bayesian networks offer potential to facilitate developing SAs, and we are developing and testing, in five case studies, a methodology for this purpose, referred to as "Control Point Bayesian Networks" (CP-BNs). The steps involved are: prepare a production chain for the commodity based around control points; identify and characterise measures as potential components of an SA; develop BNs for the SA options; analyse SA options; report. Here we provide an update on progress with the case studies and the methodology. more >

The search for a useful explanatory model based on Bayesian Networks (BNs) has now been widely established. However, such methods, in their simplest form, do not search over context-specific symmetries within the system. We introduce a new and more general class of graphical models, the Chain Event Graphs (CEGs). These capture the discrete BN as a particular subclass and further allow for asymmetries within the dependence structure of the variables. CEGs are derived from a probability tree, by merging the nodes of the tree whose associated conditional probabilities are the same, and hence retain the paths of the tree in a more compact graph. Therefore they share with the BN the property of providing an expressive graphical framework through which conclusions can be read back to the client.

We illustrate the improvement via the CEG by applying it to a subset of the Christchurch Health and Development Study looking at the effect of social and family factors on children's health in a New Zealand birth cohort. We show how we can take the MAP model of a BN search and embellish the initial analysis by a second search over CEGs. Scoring methods are set up analogously to the standard Bayesian Dirichlet metric for BNs and hence a direct comparison between the models is feasible. Our results demonstrate that the CEG search leads to a significantly higher scoring model and further that it lets us draw plausible additional conclusions from the health study. more >

All situations in which we query a BN under various scenarios deal with inference about finite populations. In this paper we reconsider the issue of finding confidence or credible intervals for queries in a Bayesian Network for finite populations. We focus on the situation in which the BN is based on discrete nodes and compare the asymptotic approaches proposed by Van Allen et al (2001, 2008) with the simulation-based approaches proposed by Donald et al (2009). A further alternative based on a single simulate of the BN is considered, namely the computation of exact binomial confidence intervals. Based on an investigation of two BN structures considered by Van Allen et al and Donald et al, we suggest that all querying of a BN should produce a probability embedded in a confidence/ credible interval, and that where the expected number of samples satisfying the particular condition queried is small, the preferred method would be the simulation method of Donald et al. (2009). However, simply computing a binomial confidence interval for a query would assist users of BNs to assess the BN and the import of conclusions drawn from it, whether using the BN for model prediction, evaluation, or validation. more >

Bayesian network (BN) software commonly represents conditional probabilities using discrete distributions. The advantage of discrete distributions is they do not assume a particular parametric form and admit to simple computations. However if observed variables are continuous quantities then they need to be discretised; but this may reduce their effectiveness in a model. That is to say that a poorly chosen discretisation can reduce the resolving power for computing posterior probabilities. A number of univariate methods are available to optimise the information represented by states beyond an equal-interval discretisation; these include quantile, minimum variance and information entropy minimisation approaches. The performance of these methods for data fitting may be evaluated by computing error rates with respect to the continuous data. But they are rarely evaluated for performance within the BN model.

This paper evaluates the performance of univariate and multivariate discretisation methods for a BN model of landslide risk. A BN model is developed from spatial data on land slide occurrences, geology, slope and drainage characteristics. The last two variables are continuous and need to be discretised for the model. We tested univariate discretisation methods with different numbers of interval states to find the best fit to historical landslide occurrence data. While this produces reasonable results, it does not consider the combination of multiple variables in determining landslide risks. A multivariate method was developed to find a discretisation that provided the highest likelihood of fitting the data. It was implemented with a Monte Carlo technique that maximised the log likelihood of a model for the historical data. The discretisation produced a lower error rate and better sensitivity rating for the Bayesian network. The method is implemented in a GIS and is intended to support multivariate discretisation for spatial variables. more >

The Australian Bureau of Meteorology is responsible for providing terminal aerodrome forecasts (TAF) that are used by the aviation industry for flight operations. Fog and low cloud forecasting is of prime importance because if fog or low cloud is forecast on the TAF aircraft must carry enough fuel to reach an alternate airport or to maintain a holding pattern above the airport until the expected clearance time. While unforecast fog/low cloud has the potential for loss of life, significant disruption and high cost, a false alarm also has economic costs because of the need to carry unnecessary fuel. Forecasting fog/low cloud is very difficult because there are many sources of uncertainty. Current forecasting guidance, such as Numerical Weather Prediction (NWP) models, have not always satisfactorily dealt with fog/low cloud and do not deal with uncertainty explicitly. Bayesian Networks, which are based on probability theory, and their extension to Bayesian Decision Networks, based on utility theory, are widely accepted as intuitively appealing, practical representations of knowledge that can be used for reasoning under uncertainty.

We developed a design strategy for building location and time specific Bayesian Networks that integrates: 1) available observations that are defined for the specific location, 2) available specific existing guidance, 3) expert opinion to overcome the limitation of a rare event and dependency on available data. We defined a set of criteria to evaluate the networks.

This design strategy was used to develop multiple networks, for different locations and lead times, depending on locations' requirements and constraints. Nineteen networks were developed for four locations for different lead times. Networks are at different stages of development, and evaluation, some have been in operational use since 2005. This approach outperforms existing guidance in a range of forecast applications, and allows forecasters to readily interpret and add value to their forecasts. more >

The design and implementation of effective environmental policies need to be informed by a holistic understanding of the system processes (biophysical, social, and economic), their complex interactions, and how they respond to various changes. Models, integrating different system processes into a unified framework, are seen as useful tools to help analyse alternatives with stakeholders, quantitatively assess their outcomes, and communicate results. This paper reviews five common approaches or model types that have the capacity to integrate knowledge (sources and types) to develop models that can accommodate multiple issues, values, scales and uncertainty considerations, as well as facilitate stakeholder engagement. The approaches considered are: systems dynamics, Bayesian networks, coupling of component models, agent-based models and knowledge-based models. We start by discussing several considerations in model development, such as the purpose of model building, the availability of qualitative versus quantitative data for model specification, the level of spatio-temporal detail required, and treatment of uncertainty. These considerations and a review of applications are then used to develop a framework to assist modellers and model users in the choice of an appropriate modelling approach for their integrated assessment applications. more >

Expansion of willows in the naturally mixed landscape of vegetation types in the Upper St. Johns River Basin in Florida, USA, impacts upon biodiversity, aesthetic and recreational values. Managers need an integrated knowledge base to support decisions on where, when and how to control willows. Modelling the spread of willows over space and time requires spatially explicit data on willow occupancy, an understanding of dispersal mechanisms and how the various life-history stages of willows respond to environmental factors and management actions. We describe an architecture for a management tool (see Fig.3) that integrates environmental spatial data from GIS, dispersal dynamics from a process model and Bayesian Networks (BNs) for modelling the influence of environmental and management actions on the key life-history stages of willows. The temporal changes in willow stages (shown in Fig.1) are modelled using a form of Dynamic Bayesian Network (DBN). Starting from a state-transition (ST) model of the willow's lifecyle, from germination to seed-producing adult, we describe the expert elicitation process used to develop a ST-DBN structure (in Netica), that follows the template described by Nicholson and Flores (2011) (see Fig.2). The ST-DBN is then converted into an Object Oriented BN (OOBN) (in Hugin) and integrated with a GIS, allowing us to model the impact over time of different management strategies on the whole region. more >

Fire danger rating systems have been developed in many fire-prone regions around the world to assist authorities in a variety of fire management activities such as assessing the potential for fires and issuing fire warning. In Australia, the McArthur's Fire Danger Rating System is operationally used to assess the potential for fires to ignite, spread and affect human property. However, this system is based only on weather parameters (i.e., wind, relative humidity, temperature and precipitation) and does not account for other critical variables (e.g., fuel load, fuel type, topography) which can have a significant influence on the potential for fire to spread and affect human property. In this paper we propose an innovative Bayesian network (BN) framework for the development of a "consequence-based" fire danger rating system. The proposed framework integrates a wide range of environmental variables (e.g., fuel type, topography, house density), models fundamental processes that govern the behaviour of fire (e.g., fire ignition and spread) and is capable of predicting the probability of property loss from fire. This BN framework can be applied in a spatially-explicit manner. The components of the BN can be linked to GIS data and generate continuous surfaces representing the probability of property loss across the landscape. The spatial-modelling facilities used to integrate BN and GIS will be discussed together with the preliminary results of its application to a case study in south-eastern Australia. more >