ICS-171:Notes 8: 1 Notes 8: Uncertainty, Probability and Optimal Decision-Making ICS 171, Winter 2001

ICS-171:Notes 8: 2 Outline Autonomous Agents –need to be able to handle uncertainty Probability as a tool for uncertainty –basic principles Decision-Marking and Uncertainty –optimal decision-making principle of maximum expected utility

ICS-171:Notes 8: 3 Autonomous Agents Consider an agent which is reasoning, planning, making decisions –e.g. A robot which drives a vehicle on the freeway Real World Current “World Model” List of possible Actions Background Knowledge Sensors Effectors Reasoning and Decision Making Agent or Robot Goals

ICS-171:Notes 8: 4 How an Agent Operates Basic Cycle –use sensors to sense the environment –update the world model –reason about the world (infer new facts) –update plan on how to reach goal –make decision on next action –use effectors to implement action Basic cycle is repeated until goal is reached

ICS-171:Notes 8: 5 Example of an Autonomous Agent A robot which drives a vehicle on the freeway Freeway Environment Model of: vehicle location freeway status road conditions Actions accelerate steer slow down Prior Knowledge: physics of movement rules of the road Sensors: Camera Microphone Tachometer Engine Status Temperature Effectors: Engine control Brakes Steering Camera Pointing Reasoning and Decision Making Driving Agent Goal: drive to Seattle

ICS-171:Notes 8: 6 The Agent’s World Model World Model = internal representation of the external world –combines background knowledge current inputs Necessarily, the world model is a simplification –e.g. in driving we cannot represent every detail every pebble on the road? details of every person in every other vehicle in sight? A useful model is the State Space model –represent the world as a set of discrete states e.g., variables = {Rainy, Windy, Temperature,.....} – state = {rain=T, windy=T, Temperature = cold,.....} An agent must –1. figure out what state the world is in –2. figure out how to get from the current state to the goal

ICS-171:Notes 8: 7 Uncertainty in the World Model The agent can never be completely certain about the external world state. i.e., there is ambiguity and uncertainty Why? –sensors have limited precision e.g., camera has only so many pixels to capture an image –sensors have limited accuracy e.g., tachometer’s estimate of velocity is approximate –there are hidden variables that sensors can’t “see” e.g., large truck behind vehicle e.g., storm clouds approaching –the future is unknown, uncertain: i.e., we cannot foresee all possible future events which may happen In general, our brain functions this way too: –we have a limited perception of the real-world

ICS-171:Notes 8: 8 Rules and Uncertainty Say we have a rule if toothache then problem = cavity But not all patients have toothaches because of cavities (although perhaps most do) So we could set up rules like if toothache and not(gum disease) and not(filling) and then problem = cavity This gets very complicated! a better method would be to say if toothache then problem = cavity with probability 0.8 or p(cavity | toothache) = 0.8

ICS-171:Notes 8: 9 Example of Uncertainty Say we have a camera and vision system which can estimate the curvature of the road ahead: There is uncertainty about which way the road is curving –limited pixel resolution, noise in image –algorithm for “road detection” is not perfect We can represent this uncertainty with a simple probability model Probability of an event = a measure of agent’s belief in the event given the evidence E –e.g., p(road curves to left | E) = 0.6 p(road goes straight | E) = 0.3 p(road curves to right | E) = 0.1

ICS-171:Notes 8: 10 Variables: Notation Consider a variable, e.g., A, –(usually in capitals) –assume A is discrete-valued –takes values in a domain e.g. binary: domain = {true, false} e.g., multivalued: domain = {clear, partly cloudy, all cloud} –variable takes one and only one value at a given time i.e., values are mutually exclusive and exhaustive –The statement “A takes value a”, or “A = a”, is an event or proposition this proposition can be true or false in real-world –An agent’s uncertainty is represented by p(A = a) this is the agent’s belief that variable A takes value a (i.e., “world is in state a”), given no other information relating to A –Basic property:  p(a) = p(A=a1) + p(A=a2) +... p(A=ak) = 1

ICS-171:Notes 8: 11 Variables and Probability Distributions –Example: Variable = Sky –takes values in {clear, partly cloudy, all cloud} –probabilities are p(clear), p(partly cloudy), p(all cloud), e.g: p(Sky = clear) = 0.6 p(Sky = partly cloudy) = 0.3 p(Sky = all cloud) = 0.1 –Notation: we may use p(clear) as shorthand for p(Sky = clear) If S is a variable, with taking values in {s1, s2, sk} –then s represents some value for S –i.e., if S is Sky, then p(s) means any of p(clear), p(all cloud), etc: this is a notational convenience Probability distribution on S taking values in {s1, s2, sk}: – P(S) = the set of values {p(s1), p(s2), p(sk)} –If S takes k values, then P(S) is a set of k probabilities

ICS-171:Notes 8: 12 Conjunctions of Events, Joint Variables A = a is an event –A is a variable, a is some value for A We can generalize to speak of conjunctions of events –A = a AND B = b(like propositional logic) We can assign probabilities to these conjunctions –p(A = a AND B = b) –This is called a joint probability on the event A=a AND B=b –Notation watch! convention: use p(a, b) as shorthand for p(A = a AND B = b) Joint Distributions –let A, B, C be variables each taking k values –then P(A, B, C) is the joint distribution for A, B, and C –how many values are there in this joint distribution?

ICS-171:Notes 8: 13 Summary of Notation and Conventions Capitals denote variables, e.g., A, B, C... –these are attributes in our world model Lower-case denotes values of variables, e.g., a, b, c, –these are possible states of the world The statement A=a is an event (equivalent to a proposition) –true or false in the real-world We can generalize to conjunctions of events –e.g., A=a, B=b, C = c (shorthand for AND(A=A, B=b, C=c) lower case “p” denotes a single probability for a particular event –e.g., p(A = a, B=b) upper case “P” denotes a distribution for the full set of possible events (all possible variable-value pairs) –e.g. P(A, B) = {p(a1,b1), p(a2,b1), p(a1,b2), p(a2,b2)} –often represented as a table

ICS-171:Notes 8: 14 Axioms of Probability What are the rules which govern the assignment of probabilities to events? Basic Axioms of Probability –1. 0 <= p(a) <= 1 probabilities are between 0 and 1 –2. p(T) = 1, p(F) = 0 if we believe something is absolutely true we give it probability 1 –3. p(not(a)) = 1 - p(a) our belief in not(a) must be “one minus our belief in a” –4. p(a or b) = p(a) + p(b) - p(a and b) probability of 2 states is their sum minus their “intersection” e.g., consider a = sunny and b = breezy One can show that these rules are necessary if an agent is to behave rationally

ICS-171:Notes 8: 15 More on Joint Probabilities and Joint Distributions “Joint” Probability = probability of conjunction of basic events –e.g., a = “raining” –e.g., b = “cold” then p(a and b) = p(a,b) = p(raining and cold) Joint Probability Distributions Let A and B be 2 different random variables –say each can take k values The joint probability distribution for p(A and B) is a table of k by k numbers, i.e., it is specified by k x k probabilities Can think of “A and B” as a composite variable taking k 2 values note: the sum of all the k 2 probabilities must be 1: why?

ICS-171:Notes 8: 16 Example of a Joint Probability Distribution We have 2 random variables –Rain takes values in {rain, no rain} –Wind takes values in {windy, breezy, no wind} rain no rain windy breezy no wind This is the table of joint probabilities Note: the sum over all possible pairs of events = 1 what is p(windy)? what is p(no rain)? Rain Variable Wind Variable

ICS-171:Notes 8: 17 Using Joint Probabilities Principle: –given a joint probability distribution P(A, B, C,...) one can directly calculate any probability of interest from this table, e.g, p(a1) –how does this work? Law of “Summing out Variables” –p(a1) = p(a1,b1) + p(a1, b2) p(a1, bk) i.e., we “sum out” the variables we are not interested in –e.g., p(no rain) = p(no rain, windy) + p(no rain, breezy) + p(no rain, no wind) = = 0.7 So, joint probabilities contain all the information of relevance rain no rain windy breezy no wind

ICS-171:Notes 8: 18 Conditional Probability Define p(a | e) as the probability of a being true if we know that a is true, i.e., our belief in a is conditioned on e being true –the symbol “|” is taken to mean that the event on the left is conditioned on the event on the right being true. Conditional probabilities behave exactly like standard probabilities –0 <= p(a|e) <= 1 conditional probabilities are between 0 and 1 – p(not(a) | e) = 1 - p(a | e) i.e., conditional probabilities sum to 1. –we can have p(conjunction of events | e), e.g., p(a and b and c | e) is the agent’s belief in the sentence on the left conditioned on e being true. Conditional probabilities are just a more general version of “standard” probabilities

ICS-171:Notes 8: 19 Calculating Conditional Probabilities p(a | b) = p(a and b) p(b) Note that p(a|b) is not = p(b|a) in general e.g., p(carry umbrella|rain) is not equal to p(rain|carry umbrella) ! An intuitive explanation of this definition: p(a | b) = number of times a and b occur together number of times b occurs e.g., p(rain|clouds) = number of days rain and clouds occur together number of days clouds occur Definition of a conditional probability

ICS-171:Notes 8: 20 Interpretation of Conditional Probabilities 2 events a and b p(a|b) = conditional probability of “a given b” = probability of a if we assume that b is true e.g., p(rain | windy) e.g., p(road is wet | image sensor indicates road is bright) e.g., p(patient has flu | patient has headache) p(a) is the unconditional (prior) probability p(a | b) is the conditional (posterior) probablity Agent goes from p(a) initially, to p(a | b) after agent finds out b –this is a very simple form of reasoning –if p(a | b) is very different from p(a) the agent has learned alot! –e.g., p(no rain|windy) = 1.0, p(no rain) = 0.6

ICS-171:Notes 8: 21 Example with Conditional Probabilities What is the probability of no rain given windy? p(nr|w) = p(nr and w) = 0.1 = 1 p(w) 0.1 What is the probability of breezy given rain ? p(b|r) = p(b and r) = 0.2 = p(r) 0.3 rain no rain windy breezy no wind

ICS-171:Notes 8: 22 Extra Properties of Conditional Probablities Can define p(a and b | c) or p(a | b and c) etc –i.e., can calculate the conditional probability of a conjunction –or the condition (term on the right of “|”) can be a conjunction –or one can even have p(a and b | c and d), etc all are legal as long as we have p(proposition 1 | proposition 2) Properties in general are the same as “regular” probabilities – 0 <= p(a | b) <= 1 –If A is a variable, then p(ai and aj | b) = 0, for any i and j, p(a1|b) + p(a2|b) p(ak|b) = 1 Note: a conditional distribution P(A|X 1,....X n ) is a function of n+1 variables, thus has k n+1 entries (assuming all variables take k values)

ICS-171:Notes 8: 23 Conditional Probability Define p(a | e) as the probability of a being true if we know that e is true, i.e., our belief in a is conditioned on e being true –the symbol “|” is taken to mean that the event on the left is conditioned on the event on the right being true. Conditional probabilities behave exactly like standard probabilities –0 <= p(a|e) <= 1 conditional probabilities are between 0 and 1 – p(a 1 |e) + p(a 2 |e) p(a k |e) = 1 i.e., conditional probabilities sum to 1. here a 2, etc., are just specific values of A –we can have p(conjunction of events | e), e.g., p(a and b and c | e) is the agent’s belief in the sentence “a and b and c” on the left conditioned on e being true. Conditional probabilities are just a more general version of “standard” probabilities

ICS-171:Notes 8: 24 Actions and States Let S be a discrete-valued variable –i.e., V takes values in the set of states {s 1,.... s k } values are mutually exclusive and exhaustive –represents a state of the world, e.g., road = {dry, wet} Assume there exists a set of Possible Actions –e.g., in driving A = set of actions = {steer left, steer straight, steer right} The Decision Problem –what is the optimal action to take given our model of the world? Rational Agent –will want to take the best action given information about the states –e.g., given p(road straight), etc, decide on how to steer

ICS-171:Notes 8: 25 Action-State Utility Matrix u (A, s) = utility of action A when the world really is in state s –u(A, s) = the utility to the agent which would result from that action if the world were really in state s utility is usually measured in units of “negative cost” e.g., u(write check for $1000, balance = $50) = -$10 –important: the agent reasons “hypothetically” since it never really knows the state of the world exactly –for a set of actions A and a set of states S this gives a utility matrix Example –S = state_of_road, takes values in {l, s, r} this the variable whose value is uncertain (=> probabilities) –A = actions, takes values in {SL, SS, SR, Halt} –u (SL, l) = 0 –u (SR, l) = $20k –u (SL, r) = $1,000k

ICS-171:Notes 8: 26 Example of an Action-State Utility Matrix ACTION Steer Left (SL) Steer Straight (SS) Steer Right (SR) Halt (H) STATE (Curvature of Road) Left Straight Right Note: you can think of utility as the negative cost for the agent => maximing utility is equivalent to minimizing cost

ICS-171:Notes 8: 27 Expected Utilities How can the driving agent choose the best action given probabilities about the state of the world? –e.g., say p(l) = 0.2, p(s) = 0.7, p(r) = 0.1 Say we take action 1, i.e., steer left We can define the Expected Utility of this action by averaging over all possible states of the world, i.e., –expected utility (EU) = sum over states of {utility(Action, state) x p(state)} EU(Steer Left) = u(SL | l) x p(l) + u (SL | s) x p(s) + u (SL | r) x p(r) = - 0 x (- 20) x (- 1000) x 0.1 = - 114

ICS-171:Notes 8: 28 Optimal Decision Making Optimal Decision Making: –choose the action with maximum expected utility (MEU) –procedure calculate the expected utility for each action choose among the actions which has maximum expected utility The maximum expected utility strategy is the optimal strategy for an agent who must make decisions where there is uncertainty about the state of the world –assumes that the probabilities are accurate –assumes that the utilities are accurate –is a “greedy” strategy: only optimizes 1-step ahead Read Chapter 16, pages 471 to 479 for more background

ICS-171:Notes 8: 29 Example of Optimal Decision-Making Use action-state utility matrix from before State Probabilities are p(l) = 0.2, p(s) = 0.7, p(r) = 0.1 Expected_Utility(Steer Left) = 0 x (- 20) x x 0.1 = -114 Expected_Utility(Steer Straight) = - 20 x x x 0.1 = -104 Expected_Utility(Steer Right) = - 20 x x x 0.1 = -18 Expected_Utility(Halt) = x (- 500) x (-500) x 0.1 = -500 Maximum Utility Action = “Steer Right” –note that this is the least likely state of the world! –but is the one which has maximum expected utility, i.e., it is the strategy which on average will minimize cost

ICS-171:Notes 8: 30 Another Example: Renewing your Car Insurance What is the optimal decision given this information? –EU(Buy) = x ( ) = –EU(Not Buy) = 0 x (-2000) x (-90000) x 0.01 = = ACTION Buy Insurance Do not buy State probabilities STATE (accident-type in next year) NoneMinorSerious

ICS-171:Notes 8: 31 Summary Autonomous agents are involved in a cycle of –sensing –estimating the state of the world –reasoning, planning –making decisions –taking actions Probability allows the agent to represent uncertainty about the world –agent can assign probabilities to states –agent can assign utilities to action-state pairs Optimal Decision-Making = Maximum Expected Utility (MEU)Action