Innovation at IDI

Turbo Model Building for Bayesian Networks

We have developed a process for building Bayesian network models in two-day, facilitated meetings attended by analysts, model developers, and external subject-matter experts. The facilitator guides the participants through the steps in the development process (described in the following sections), elicits estimates of model parameters, and ensures that the requirements of the methodology are met. A second member of the modeling staff implements the model on the computer and takes notes. The model is projected onto a screen during the development process so that all participants are aware of the variables and relationships included in it. Both analysts and external experts provide the information and assessments that are incorporated into the model. The analysts usually provide critical information about the questions to be addressed by the model, while all participants provide the knowledge incorporated into the model. 

The facilitator’s behavior is critical in two respects. First, the technical aspects of applying Bayesian analysis must be guided by an expert – the international relations and political science educations of most analysts don’t prepare them for this methodology. The challenges of helping the group to frame questions properly – consistent with probability theory – and to keep their engagement fresh while estimating large conditional probability tables are not trivial items. In addition, the facilitator helps to keep the gate open to contrary data and judgments and healthy debate, to elicit contributions from all members, to challenge what everyone takes for granted, and to curb the natural tendencies of dominant actors to hog the stage and dictate the analysis – all while demonstrating respect for each contribution.

For more information, see the paper: It's the People, Stupid

Combining Equation-based Models and Bayesian Network Calculations via Extend and Netica

An IDI Analyst has created a library of modeling blocks for the Extend Modeling and Simulation Software package that allows Extend to dynamically interact with a Netica Bayesian network or influence diagram during discrete or mixed continuous/discrete simulation.  This combination provides some unique capabilities for both the users of Extend and Netica.  For Netica Bayesian network users, the library allows for the construction of a simulated environment in Extend that can be used to teach a Bayesian network or influence diagram the correct probabilities and utility values.  This allows an influence diagram to learn from the simulation how to make “good” decisions. This is useful if there is no source of data or experts from which to generate the probabilities and utilities for the network.  This method has been shown to be particularly useful where the inputs to the network may include errors, missing data or when data may come from an environment where denial and deception are likely.  Even if the Bayesian network or influence diagram is created through manual construction techniques, the Extend modeling environment provides a method to test the network to see how it reacts in its expected environment. 

Users of the Extend software package can also realize advantages by using the library.  The Extend package as provided by the vendor is limited to rule-based decision making.  The addition of a Bayesian network or influence diagram allows for dynamic multi-attribute decisions to be made during a discrete simulation.  This allows for much better artificial intelligence resulting in decisions within a simulation that are much closer to human decision making. Additionally, Bayesian networks have been demonstrated to represent complex, abstract entities that cannot be described by mathematical equations as probabilistic models.  This approach is much simpler than using look-up tables and has been shown to be more accurate than using simplified mathematical expressions.

For more information, see the paper: Integrating Equation Models and Bayesian Networks

Implementing Dynamic Decision Networks (DDNs)

DDNs are an ideal aid for representing (or supporting) a decision maker faced with multiple uncertain variables and many opportunities to collect information.  It is also the case that these concepts can be applied to represent multiple competing objectives in the form of multi-attribute objective function.  

DDNs are being applied to represent a decision making process within the US Army’s Future Combat System Program.  Some examples of decisions modeled during our work are shown in the table below.  These examples are meant to demonstrate the repetitive nature of these decisions.

For more information, see the briefing: DDN Overview

Modeling Adaptive Adversaries

IDI was tasked with developing a model to assess the deterrent effect of a specific U.S. Military Force on the decision making of a specific adversary in a conventional scenario. Modeling results are not presented due to classification. The emphasis was on the development process and how it might be applied generically to other U.S. Forces and adversaries. This was a two phased effort. During the first phase a prototype model was developed to assess the deterrent effect of U.S. military actions on an adversary’s tactical decision making. The second phase prototype assessed the deterrent effect of U.S. military actions integrated with strategic U.S. Diplomatic, Information, and Economic actions on an adversary’s strategic decision making over time.

For more information, see the paper: Modeling an Adversary's Decision Making Process

For an another task, IDI developed an approach to model the behavior of the “Adaptive Adversary” (Red) as acting and reacting agent in terrorism risk assessment and risk management. The plan: Within the probabilistic risk assessment (PRA) framework, we set out to populate the chance event nodes representing Red choices (weapons, targets, etc.) with a probability for each branch (each weapon form, each target, etc.) to determine the probability that Red would pick that branch. The term “Adaptive Adversary” refers in particular to an adversary who will shift his focus (among weapons, targets, etc.) in response to newly deployed defensive measures (radiation and biological agent detectors, etc.).

For more information, see the paper: Adaptive Adversary Modeling