Nova Scotia Federation of Agriculture: Risk Proofing Nova Scotia Agriculture This presentation outlines some of the basic concepts and components of a Bayesian Network – the program and approach that we are using to build the AgriRisk risk assessment tool. April 6, 2017 1
Nova Scotia Federation of Agriculture: Risk Proofing Nova Scotia Agriculture There are many definitions of Risk. For this project we use a simple one which can be expressed in plain language or as a quasi-probability statement. The middle statement illustrates that Risk in this project is seen as a conditional probability. The last statement is an example of an actual risk expressed as a conditional probability statement – this would be read as “the probability of the yield of grapes being less than our target yield is conditional on the seasonal weather forecast”. April 6, 2017 2
Nova Scotia Federation of Agriculture: Risk Proofing Nova Scotia Agriculture Risk perceptions are fundamentally important to what people do. In part this project seeks to work with key stakeholders to assist with generating more informed and reliable perceptions of risk. April 6, 2017 3
Nova Scotia Federation of Agriculture: Risk Proofing Nova Scotia Agriculture The grape and wine industry in Nova Scotia is interconnected which means that uncertainties, worries, or risks at one end of the system (or somewhere in the middle) can ripple or reverberate through the whole system. Decisions or responses associated with these risk at one part of the chain can be felt elsewhere. April 6, 2017 4
Nova Scotia Federation of Agriculture: Risk Proofing Nova Scotia Agriculture For the AgriRisk project we are using a Bayesian Network to develop a model that will help understand the impacts of risks at various places in the grape and wine value chain and the impacts that changes in one part of the chain have on the others. This slide highlights some of the key differences between a Bayesian and classical or frequentist approach to probability. The take-home message here is that the Bayesian approach uses the best evidence available at this particular time (which may be in the form of expert opinion), to understand probabilities of outcomes or events and allows us to build on this as new information and data becomes available. So we can continuously update a Bayesian model to help inform our understanding of risk (or opportunity) in the grape and wine value chain. April 6, 2017 5
Nova Scotia Federation of Agriculture: Risk Proofing Nova Scotia Agriculture Bayesian Networks have two key features: their structure or the relationships among nodes (variables) and their conditional probability tables (CPTs), which are behind the scenes, defining the conditional probabilities of each node. These CPTs are the engine of the Bayesian Network. The CPT example on the right shows the conditional probabilities of node E on the left. The right hand side of the table shows the probabilities for each state of E given every possible combination of A and B which are the parent nodes for E. In other words, E is conditional on A and B or you could say that A and B are factors that determine the state of E. Each row of probabilities in the table will sum to 1. April 6, 2017 6
Nova Scotia Federation of Agriculture: Risk Proofing Nova Scotia Agriculture This is a sample network using dummy numbers that illustrates some of the possible variables and relationships for a grape yield and profit model. Blue nodes reflect climate variables; green nodes management variables; grey nodes statistical parameters (e.g. the mean and standard deviation are statistical parameters for the normal distribution); brown nodes are biophysical variables and yellow nodes are economic variables. The numbers or text on the left hand side of each node are the possible states for that particular variable (remembering that in BNs all continuous variables like temperature, weight, or cost have to be converted into discrete variables). The bars to the right and the numbers in the central column reflect the probability associated with each state. For example looking at the large yellow box “profit” you can see there is a 31.9% chance (or probability of 0.319) that profit takes the state “2000 to 3000”. The average (and standard deviation, or variation) profit is shown at the bottom of the profit node and equals 2480 (2300). This representation is a summary of the probability structure in the data. When we have the actual network completed through the AgriRisk project, a user will be able to click on any of the states for any of the nodes and see how that selection impact the probabilities of states elsewhere in the grape value chain. An important point to note about assessing risk using a BN in this way is that they can simultaneously identify opportunities (such as the probability of profits exceeding a target level) as well as risks. April 6, 2017 7
Nova Scotia Federation of Agriculture: Risk Proofing Nova Scotia Agriculture In this slide we highlight the key advantages of BN models and also identify some things to be aware of in their use. Because of their linked nature BNs are ideal for value chain analyses; they can use virtually any form of data when building the underlying Conditional Probability Tables ~ expert knowledge, data from observations or measurements, relationships derived from research; the models can be continuously improved as new knowledge or data becomes available; they enable us to explore risks with uncertainty clearly presented; sensitivity analyses enables us to identify which variables in the model contribute most to our uncertainty in the state of any other variable; BNs are efficient when it comes to using sparse data sets. We need to be cautious however that continuous variables like temperature or cost need to be discretised (cut into discrete ranges) and BNs do not easily deal with feedbacks (feedbacks are cycles such as this year’s profit affecting next years management decisions); presentation of complex BNs can be daunting to those unfamiliar with them; like any other modelling tool they are not a silver bullet to solve any problem – they are only as good as the data and knowledge that goes into them. April 6, 2017 8
Nova Scotia Federation of Agriculture: Risk Proofing Nova Scotia Agriculture In this slide we address some common questions about BN models as we propose to apply them. Netica is a Canadian computer program that can be downloaded online here: http://www.norsys.com/ R, which is the statistical programming language behind the scenes, is open source and available here: https://www.r-project.org/ April 6, 2017 9
Recommend
More recommend
