BayesiaLab Webinar Series
Adversarial Reasoning with Bayesian Networks—
Applying the Math of Denial & Deception
Webinar on Friday, October 12, 2018, at 1 p.m. (CDT, UTC -05)
Background & Motivation
This webinar expands on the topic of systematic deception and counterdeception, which we first touched upon in our seminar on Intelligence Analysis with Bayesian Networks in Arlington, Virginia, on September 11, 2018. We now focus specifically on reasoning in the presence of a geopolitical adversary, who is prepared to deceive us, and whom we are willing to mislead, too. Concepts that are typically frowned upon, if not downright illegal in business and civil society, we will now explore systematically for purposes of national security.
The Drama of Deception
The capacity of being able to distinguish fact from fiction we often attribute to personal experience in life ("street smarts") as opposed to knowledge acquired academically. Natural language has certainly equipped us with a remarkably broad range of vocabulary capturing all nuances of deception. And, the most elaborate use of deceit we frequently observe in crime dramas, in which the process of solving the mystery is central to entertaining the audience. In such plots, it typically is brilliant intuition that leads to a resolution rather than dry mathematics, i.e., more art than science.
As we discussed in the earlier seminar, given the prevalence of biases in reasoning under uncertainty, relying on human judgment can be problematic. Once we add hostile intent and willingness to deceive into an already uncertain situation, normative reasoning appears difficult to achieve. It is easy to imagine how multiple layers of mutual deception can produce seemingly intractable complexities for analysts and decision makers.
One Word: Inference
In sharp contrast to the wide array of terms available to describe deception, the mathematical treatment is straightforward: it reduces to performing inference, i.e., updating a belief given evidence, which can be computed with the well-known Bayes' Rule. This formula can characterize both sides of the deception, i.e., the one who deceives and the one being deceived. While the simplicity of this theory is very appealing, the practical implementation remains challenging. In fact, it is virtually intractable without the aid of a computer.
Practical Bayesian Inference with Bayesian Networks
Bayesian networks are an ideal framework for representing probabilistic relationships in high-dimensional and complex problem domains. Even with extreme degrees of uncertainty, we can encode relationships between variables and formally reason with them. We can conveniently quantify uncertainty by calculating the entropy of variables. Computing Mutual Information helps us understand what kind of observations would reduce our uncertainty regarding any target variable of interest—and by how much. Furthermore, by looking at the Bayes Factor, we can see to what extent new observations are consistent with previous ones. Thus, we have a formal measure for the plausibility of what we observe about an adversary. At the same time, we know what evidence we need to produce, or rather fake, for the adversary to perceive what we wish him to believe.
For those accustomed to working with data, none of the above information-theoretic concepts are novel. What we will present in this webinar is that we can now utilize these very same concepts in domains for which little reliable data exists.
Implementation with BayesiaLab
The novel aspect of this webinar is that we translate the attractive theoretical properties of Bayesian inference and Bayesian networks into a highly practical tool for analysts and decision makers through the BayesiaLab software platform. We can take any available domain knowledge and encode it directly in the form of a directed graph. This can be based on an individual's understanding or on the collective knowledge of a group of experts. The latter topic we will explore in a separate seminar about the Bayesia Expert Knowledge Elicitation Environment in Arlington, Virginia, on November 13.
From Observation to Intervention
While much of this webinar's discussion focuses on detecting and counteracting hostile deception, a decision maker needs to take reasoning one step further. He or she needs to evaluate the potential consequences of actions not yet taken, which necessitates the formal simulation of the interventions under consideration. This prompts a formal shift from observational to causal inference. However, given that we have already encoded our problem domain as a causal Bayesian network, this very different kind of inference can be immediately performed with the same model. This means that we can simulate an enemy's deception, our counterdeception, plus our course of action concurrently in a single model.
- Motivation: Using Denial, Deception, and Counterdeception in Conflicts
- Why Bayesian Networks as a D&D Framework?
- Mathematizing Denial & Deception:
- Distributions of Evidence
- Joint Probability Distributions
- Uncertainty and Entropy
- Mutual Information
- Bayes Factor, Consistency, Conflict
- Observational Inference
- Causal Inference (Intervention)
- Synthesizing Evidence for Denial and Deception
- Likelihood Matching
- Detecting Deceit
- Discovering High-Dimensional Anomalies
- Interpreting Anomalous Observations
- Simulating Interventions
Who Should Attend?
This webinar is intended primarily for the intelligence community, members of the military, defense contractors and consultants, e.g.:
- Intelligence analysts
- J2/G2 officers
- Policy analysts
- Strategy analysts
- Knowledge managers
- Risk analysts
- Research investigators
- Decision scientists
- Operations Research analysts
- Counterterrorism analysts
- Wargaming, modeling/simulation analysts
Students and teachers in related fields are welcome to join.