Artificial Intelligence: Probabilistic Reasoning

Human intelligence is based on constant learning and past experience. We use our prior knowledge and use it in one or the other way to do things. In more or less same manner, we can have machines with intelligence. Machines (no, not those military machines) i.e. computers can be fed intelligence to assist us with various tasks. The branch of Computer Science that deals with the study of machine intelligence is termed as Artificial Intelligence and its roots are dated as early as in the year 1956.

Artificial Intelligence is useful in numerous areas. Over the period of 50 plus years many techniques and tools have been developed to solve a variety of difficult problems. Neural networks, Probabilistic methods for uncertain reasoning, Statistical learning methods and Intelligent search and optimization are some of  the tools/techniques used to solve problems in domains where Artificial Intelligence is useful. Let's understand how Probabilistic methods for uncertain reasoning works. 

Consider for example, an application where we want to have a system that can help us diagnose a disease for a given patient (This is called as Probabilistic Inference). The end result would certainly be a probability value. The system will more or less acts as a Doctor. Now, what set of inputs does this system take? We might have inputs like different symptoms patient has, results of medical test(s) and possibly patient's habits (e.g drinking, smoking etc). The set of inputs really depend upon the type of diagnostic system we want to develop. The set of  inputs are expressed in terms of probability values. The system being developed goes through 3 phases: Representation, Inference and Learning.

Each of these individual inputs are related to other inputs and they kind of influence one another. The interaction of all these inputs is captured via a graphical model, which is constructed using either a Bayesian network or a Markov network - Representation. Each node in the graphical model corresponds to one of the inputs. So for example, when we have that patient is a chain smoker, the probability that patient exposes a chance of a certain type of cancer goes up - Inference. So, using the information available to us (information fed to the system), we eventually come to a conclusion based on the interaction we have in our graphical model.

But, before we can do inference from the system we have to train the system with some prior data - Learning. This type of data is termed as the training set. This training data helps us understand correlations between inputs . This kind of system helps us deal with uncertainty that lies within these type of applications and gives us results in terms of probabilities. 

Few other examples where these technique is useful are: Robot navigation, Text analysis, Speech recognition (we answer probabilistic-ally about what word it might be) and Gene regulatory networks.  I am working in the same area (called Probabilistic Graphical Model) for my master's project (CS 298).