Last Updated on
This post is a spotlight interview with Jhonatan de Souza Oliveira on the topic of Bayesian Networks.
Could you please introduce yourself?
My name is Jhonatan Oliveira and I am an undergraduate student in Electrical Engineering at the Federal University of Vicosa, Brazil. I have been interested in Artificial Intelligence since the beginning of college, when had my first adventure investigating and building a simple chatbot for a Symposium website. I also am a member of an autonomous robot soccer team called “BDP – Believe, Do and Play”, where we research and develop technologies for the RoboCup category called “Small Size League”.
In late 2012, I got a scholarship for a exchange program through a Brazilian government program called “Science Without Borders“. I went to University of Regina, where I could have the pleasure of meet Dr. Cory Butz, a well known Bayesian Network researcher.
Since then, I have been developing research in Bayesian Networks inference and modeling with Dr. Butz and a Brazilian friend, Andre Evaristo.
What are Bayesian Networks?
In general, Bayesian Networks (BNs) is a framework for reasoning under uncertainty using probabilities. More formally, a BN is defined as a Directed Acyclic Graph (DAG) and a set of Conditional Probability Tables (CPTs). In practice, a problem domain is initially modeled as a DAG. Lets take an example from the good reference Bayesian Networks Without Tears (PDF):
Suppose when I go home at night, I want to know if my family is home before I open the doors. From my knowledge, I can model a DAG with the following information: usually, when my family is out, the outdoor light is on and the dog is out, but the dog is also out when it has bowel problems. If the dog is out I can hear its bark. This is the DAG for this problem, where the arrows indicate a causal idea:
From the probability side, we would need a Joint Probability Distribution (JPD) – in practice, for finite and discrete variables it would be a big table where each row associate a configuration of the variables’ domain to a probability value. In our example, all 5 variables are binary and the JPD would have 2^5 = 32 rows, which may not sound now a big deal but in more complex problems it can be intractable.
The intractability of the JPD was solved by Judea Pearl in a series of papers published since 1988. He proposed the use of conditional independences to break that big JPD table into small ones (yes the CPTs).
In our example, we only need to provide 5 small tables, which in total will sum to 2^2 + 2^1 + 2^3 + 2^1 + 2^2 = 20 rows.
After the problem domain is modelled with a DAG and a CPTs, we want to run some inference algorithm in order to update the model after absorbing some new evidence, say the light is on; or even answer some query about the problem, say what is the probability of hear a bark given that the light is on?
Why do we need Bayesian Networks?
Humans are not good with reasoning in systems with limited or conflicting information. Consider a web search engine where the user type in a query and the system provides a list of results. Which web page is more relevant to this specific user? Now, consider a medical diagnosis system, in which a patient has some, but not all, of the symptoms of a disease.
It would be handy if we have something to manage all this limited/conflicting information. So, here is why we need them: BN is a framework for uncertainty management.
What are some popular examples where Bayesian Networks were used?
BNs has been used in many fields, including improvements on propulsion systems, maintenance system and diagnostic analysis, among many others.
The paper “Display of Information for Time-Critical Decision Making” shows how NASA used BNs to manage “the complexity of information displayed to people responsible for making high-stakes, time critical decisions“.
Microsoft have used BNs to build a assistance software able to help users when they get stuck using Windows. The Lumiere Project, as it’s called, was used in the Office Assistant in Microsoft Office ’97 suite and was an intelligent user interface able to model a problem domain given the user’s background, actions and queries.
The heart disease program at MIT has as goal “to assist physicians in the diagnosis of patients with cardiac symptoms, focusing on hemodynamic dysfunction“. BNs are used to update the beliefs’ given the symptoms, and then inference is used to get the most probable diagnosis.
Why aren’t Fuzzy Logic and Rule based Systems good enough reasoning systems?
Fuzzy Logic is not always intuitive. For example, consider flipping a fair coin, where it lands on heads has 0.5 as an assigned value and lands on tails has 0.5. Now, what is the belief that a given coin flip will land on heads “or” tails? According to fuzzy logic, two predicates related to an “or” operator has as result the minimum of both predicates assigned value, in our example, it would be 0.5 or 50%. Although we are 100% sure that the coin will land on heads “or” tails.
On the other hand, ruled based systems have a lack of semantics. A confidence-factor is used to measure the uncertainty in a set of “if-then” structures. For instance, consider a case where “If the website does not open then I lost my internet connection – with a coefficient of 9“. One could ask why the coefficient 9, or why not 19 or 900, or even ask if 9 is high or low. Yet, ruled based systems can become exponentially big trying to capture all conditions in a domain problem, in some cases becoming infeasible. For instance, if the website does not open my belief in “lost connection” becomes high, although we are not checking the possibility of a server problem or even a browser error.
What are some limitations with Bayesian Networks?
Probably the most notable weakness of BNs is the designing methodology.There is no standard way of building BNs.
The design of a BN can be a considerable amount of effort in complex systems and it is based on the knowledge of the expert(s) who designed it. Although, this disadvantage can be good in another point of view, since BNs can be easily inspected by the designers and has the guarantee that the domain specific information is being used.
What resources would you recommend to a beginner in Bayesian Networks?
For a beginner in BN (but with some AI knowledge), I would start with the excellent Bayesian Networks without Tears (PDF).
For textbooks on Bayesian Networks, I recommend:
- Probabilistic Graphical Models: Principles and Techniques
- Expert Systems and Probabilistic Network Models
- Modeling and Reasoning with Bayesian Networks
- Introduction to Bayesian Networks
An excellent academic resource is the Association for Uncertainty in Artificial Intelligence (AUAI). And, of course, Judea Pearl website is a rich resource for BNs stuff.