advertisement

Divya Aggarwal Cis 203 Final Project Reasoning with Uncertainty Probability and Bayes Networks Consider this scenario. You are debugging code in a programming language like java or prolog. All of a sudden, the input/output display screen refuses to respond. According to you, this can result from either the hardware not responding or there is a bug in your code. If the hardware is not responding then the other applications shouldn’t run as well. But this may not be the case. There can be a bug in your code. But your code has been tested and run on other systems and runs perfectly. So the other plausible explanation can be the interpreter is running slowly, your processor may be slow. Thus, you can say that “most of the times, the interpreter runs well with the code bug free and the hardware not malfunctioning.” You may even give a percentage to your “most of the times”. You may say, “seventy percent of the times the interpreter runs well with the code bug free and the hardware not malfunctioning.”( Artificial Intelligence: Theory and Practice, Thomas Dean, James Allen, Yiannis Aloimonos.) The ability to reason with uncertain information is an indispensable requirement for modeling intelligent behavior in a complex and dynamic environment. This is why uncertain reasoning has become a major research topic of AI with many important applications. The challenges of cognitive robotics and intelligent agent systems have recently increased the demand for appropriate, powerful and efficient formalisms and methodologies to deal with uncertainty. Classical inferential models do not permit the introduction of prior knowledge into the calculations. For the rigours of the scientific method, this is an appropriate response to prevent the introduction of extraneous data that might skew the experimental results. However, there are times when the use of prior knowledge would be useful to the evaluation process. Probability theory allows us to make qualitative as well as quantitative claims. There are many propositions whose truth-values you do not know prior to having to make a decision that depends on them. Various formalisms have been employed to represent and reason about uncertain information, example, MYCIN and PROSPECTOR. The formalism that is most well developed is based on probabilities. In ordinary logic we can deduce Q from P and P > Q. That is if an agent knows P > Q, and it subsequently learns P, it can infer Q also. Assuming that we have a collection of random variables, V1, V2,…,Vk and if we want to talk about the value of Vi without saying what the value is, we use the symbol vi Probability functions must satisfy certain properties: a.) 0 <= p(V1,V2,…..Vk) <= 1 b.) (p(V1,V2,….Vk) = 1 Generally we are concerned with a set of random variables. We denote the joint probability that the values of V1,V2…Vk are v1,v2….vk by the expression p(V1 = v1, V2 = v2,……Vk = vk). This expression is the joint probability function over the variables V1,…..Vk. It maps the set of variables in to a real number between 0 and 1. When we know the values of all of the joint probabilities for a set of random variables, we compute the marginal probability of one of these variables.(Artificial Intelligence: A new synthesis, Nilsson). At times we want to use information about the values of some variables to obtain probabilities for the values of others. Almost all probability statements involve implicit or explicit conditions. For example, a doctor might assign a probability to a patient having measles. Also the doctor has made an implicit assumption that the patient is a person and not an animal. Hence let M be a random variable representing the proposition that the patient has measles and P be a random variable representing the proposition that the patient is a person. Let Measles be the set of all patients with measles and People be the set of all people. Hence the probability that the next patient entering the doctor’s office has measles is P(M True | P True) | Measles People | | Measles | | People | | People | The conditional probability function of Vi given Vj is denoted by P(Vi | Vj). For any values of the variables Vi and Vj, it is given by p(Vi | Vj) p(Vi, Vj) p(Vj) where P(Vi. Vj) is the joint probability of Vi and Vj and P(Vj) is the marginal probability of Vj. A conditional probability is a normalized version of a joint probability. We can also express a joint probability in terms of a chain of conditional probabilities. The general form for this chain rule is k P(V 1,V 2,....Vk ) P(Vi | Vi 1,...,V 1) i 1 Since the way in which we order variables in a joint probability function is unimportant, we can write P(Vi, Vj) = p(Vi | Vj) p (Vj) = p(Vj | Vi) p(Vi) = p(Vj, Vi). Hence, P(Vi | Vj) P(Vj | Vi) P(Vi) P (Vj) This is the Bayes rule. A variable V is conditionally independent of a set of variables Vi, given a set Vj, if p(V | Vi, Vj) = p(V | Vj) and use the notation I(V | Vi, Vj). The reason behind consitional independence is that if I(V | Vi, Vj ), then Vi tells us nothing more about V than we already knew by knowing Vj. A single variable Vi is conditionally independent of another variable Vj, given a set V we have P (Vi | Vj, V) = P(Vi | V ). By definition of conditional probability, we have P (Vi | Vj, V)P(Vj|V) = P(Vi,Vj | V ). As a generalization of pairwise independence, we say that variables V1….Vk are mutually conditionally independent. In a set V if each variable is conditionally independent of all the others, given V . Hence, k P(V 1, V 2...Vk | V ) P(Vi | Vi 1,...V 1,V ) i 1 k P(V 1, V 2...Vk | V ) P (vi | V ) i 1 Conditional independencies can be conveniently represented in structures called Bayes Networks.(Artificial Intelligence : A new synthesis, Nilsson) Bayes' Theorem, developed by the Rev. Thomas Bayes, an 18th century mathematician and theologian, was first published in 1763. Mathematically it is expressed as : P( H | E , c) P( H | c) P( E | H , c) P( E | c) where H is the hypothesis given the additional evidence E and the background context c. The left hand term P (H|E,c) is the probability of H after considering the effect of E on c. The term P (H|c) is called the prior probability of H given c alone. The term P (E|H,c) is called the likelihood" and gives the probability of the evidence assuming the hypothesis H and the background information c is true. Finally, the last term P (Elc) is independent of Hand can be regarded as a normalizing or scaling factor. Bayes Networks: A Bayesian Network is a Directed Acyclic Graph where nodes represent assertions and where an arc from a node A to a node B expresses that A is a cause of B. For each arc (A B) we are given a "link" matrix expressing the probability P(B|A), that specifies for each value of A the probability of each value of B.(yoda.cis.temple.edu) For example, if A is Age, with values (Young, Middle, Old) and B is Income, with values (Low, Medium, High), then P (Income| Age) is of the form Income AGE Young Middle Old Low 0.90 0.60 0.65 Medium 0.08 0.30 0.25 High 0.02 0.10 0.10 Definition: Consider a domain of n variables, x1….xn. each variable may be discrete having a finite or countable number of states or continuous. Given a subset X of variables xi where xi U. An instance (state of every variable in x) of X is denoted as X = P(xy|x1…xy-1, ) = P(xy|y, )kx. (www.gpfn.sk.ca~daryle/papers/bayesian_networks) As an example of Bayesian Networks, given a situation where it might rain today and might rain tomorrow, what is the probability that it will rain on both days. Rain on two consecutive days is not independent events with isolated probabilities. If it rains the first day, then it is more likely to rain the next. Thus determining the chances that it will rain today and then determining that it will rain tomorrow conditional on the probability that it will rain today is called determining joint probabilities. The probability of such joint events is expressed as P(E1,E2) = P(E1)P(E2|E1) Expressing the joint probabilities in a table format: Marginal and Joint Probabilities for rain both today and tomorrow Rain Tomorrow No Rain Tomorrow Marginal Probability of rain today Rain today 0.14 0.06 0.20 No rain today 0.08 0.72 0.80 Marginal 0.22 0.78 Probability of rain tomorrow The same scenario can be explained using a Bayesian Network Diagram. P(E1,E2) = 0.14 P(E1,!E2) = 0.06 P(!E1,E2) = 0.08 P(!E1,!E2) = 0.72 (www.gpfn.sk.ca~daryle/papers/bayesian_networks) In Bayesian Networks only one branch needs to be traversed. Instead of using four joint probabilities we can use the independence of the parameters to limit our calculations to two. Friedman and Goldszmidt suggest this example to explain Bayesian Networks. Consider 5 random variables: “Burglary”, “Earthquake”, “Alarm”, “Neighbor’s Call”, and “Radio Announcement”. Here the independent variables are “Burglary” and “Earthquake”. “Burglary” and “Radio Announcement” are independent given “Earthquake”. This means that although a radio announcement will result from an earthquake, it won’t from a burglary. Because of the independence among these variables the probability of P (A, R, E, B) is : P (A, R, E, B) = P(A|E,B)*P(R|E)*P(E)*P(B) Expressing this as a Bayesian network. Earthquake Burglary Alarm Radio Announcement Neighbor’s Call Advantages of using Bayesian Networks: Independence among variables is easy to recognize and isolate while conditional relationships are delimited by a directed graph edge: two variables are independent if all the paths between them are blocked. Not all joint probabilities need to be calculated to make a decision. Uses and applications of Bayesian Networks: 1. NASA’s Ames Research Center is interested in deep space exploration and knowledge acquisition. The AutoClass project is an attempt to create Bayesian applications that can automatically interpolate raw data from interplanetary probes and deep space explorations. 2. Lumiere: Lumiere is software that can automatically and intelligently interact with software users by anticipating their goals and needs. 3. They are also used in the answer wizard of Office 95 and Office Assistant of Office 97. 4. Vista: This is a decision-theoretic system, developed by Eric Horvitz, in use at the NASA Mission Control Center in Houston. It uses Bayesian Networks to interpret live telemetry and provide advice on the likelihood of alternative failures on the space shuttle’s propulsion systems. 5. Bayesian Networks are also used independently in different communities: genetics (linkage analysis), speech recognition, tracking (Kalman Filtering), data compression (density estimation) and coding (turbo codes). 6. Nice Argument Generator: this uses two Bayesian networks to generate and assess natural language arguments. Thus probabilistic inference in Bayes networks dominates other methods, like fuzzy logic and possibility theory, for most expert system applications, but the subject remains controversial. Bibliography: 1. http://www.gpfn.sk.ca/~daryle/papers/bayesian_networks/bayes.html 2. Artificial Intelligence: A new Synthesis, Nilsson, 1998 3. Artificial Intelligence: Theory and Practice, Thomas Dean, James Allen, Yiannis Aloimonos.