QUESTION BANK. SUBJECT CODE / Name: CS2351 ARTIFICIAL INTELLIGENCE UNIT V. PART -A (2 Marks) LEARNING

Transcription

1 QUESTIO BAK DEPARTMET: CSE SEMESTER: VI SUBJECT CODE / ame: CS2351 ARTIFICIAL ITELLIGECE UIT V PART -A (2 Marks) LEARIG 1. Explain the concept of learning from example. Each person will interpret a piece of information according to their level of understanding and their own way of interpreting things. 2. What is meant by learning? Learning is a goal-directed process of a system that improves the knowledge or the Knowledge representation of the system by exploring experience and prior knowledge. 3. How statistical learning method differs from reinforcement learning method? Reinforcement learning is learning what to do--how to map situations to actions--so as to maximize a numerical reward signal. The learner is not told which actions to take, as in most forms of machine learning, but instead must discover which actions yield the most reward by trying them. In the most interesting and challenging cases, actions may affect not only the immediate reward but also the next situation and, through that, all subsequent rewards. These two characteristics--trial-and-error search and delayed reward--are the two most important distinguishing features of reinforcement learning. 4. Define informational equivalence and computational equivalence. A transformation from on representation to another causes no loss of information; they can be constructed from each other. The same information and the same inferences are achieved with the same amount of effort. 5. Define knowledge acquisition and skill refinement. knowledge acquisition (example: learning physics) learning new symbolic information coupled with the ability to apply that information in an effective manner skill refinement (example: riding a bicycle, playing the piano) occurs at a subconscious level by virtue of repeated practice 6. What is Explanation-Based Learning? The background knowledge is sufficient to explain the hypothesis of Explanation- Based Learning. The agent does not learn anything factually new from the instance. It extracts

2 general rules from single examples by explaining the examples and generalizing the explanation. 7. Define Knowledge-Based Inductive Learning. Knowledge-Based Inductive Learning finds inductive hypotheses that explain set of observations with the help of background knowledge. 8. What is truth preserving? An inference algorithm that derives only entailed sentences is called sound or truth preserving. 9. Define Inductive learning. How the performance of inductive learning algorithms can be measured? Learning a function from examples of its inputs and outputs is called inductive learning. It is measured by their learning curve, which shows the prediction accuracy as a function of the number of observed examples. 10. List the advantages of Decision Trees The advantages of Decision Trees are, It is one of the simplest and successful forms of learning algorithm. It serves as a good introduction to the area of inductive learning and is easy to implement. 11. What is the function of Decision Trees? A decision tree takes as input an object or situation by a set of properties, and outputs a yes/no decision. Decision tree represents Boolean functions. 12. List some of the practical uses of decision tree learning. Some of the practical uses of decision tree learning are, Designing oil platform equipment Learning to fly 13.What is the task of reinforcement learning? The task of reinforcement learning is to use rewards to learn a successful agent function. 14. Define Passive learner and Active learner. A passive learner watches the world going by, and tries to learn the utility of being in various states. An active learner acts using the learned information, and can use its problem generator to suggest explorations of unknown portions of the environment.

3 15. State the factors that play a role in the design of a learning system. The factors that play a role in the design of a learning system are, Learning element Performance element Critic Problem generator 16. What is memorization? Memorization is used to speed up programs by saving the results of computation. The basic idea is to accumulate a database of input/output pairs; when the function is called, it first checks the database to see if it can avoid solving the problem from scratch. 17. Define Q-Learning. The agent learns an action-value function giving the expected utility of taking a given action in a given state. This is called Q-Learning. 18. Define supervised learning & unsupervised learning. Any situation in which both inputs and outputs of a component can be perceived is called supervised learning. Learning when there is no hint at all about the correct outputs is called unsupervised learning. 19. Define Bayesian learning. Bayesian learning simply calculates the probability of each hypothesis, given the data, and makes predictions on that basis. That is, the predictions are made by using all the hypotheses, weighted by their probabilities, rather than by using just a single best hypothesis. 20. What is utility-based agent? A utility-based agent learns a utility function on states and uses it to select actions that maximize the expected outcome utility. 21. What is reinforcement learning? Reinforcement learning refers to a class of problems in machine learning which postulate an agent exploring an environment in which the agent perceives its current state and takes actions. The environment, in return, provides a reward (which can be positive or negative). Reinforcement learning algorithms attempt to find a policy for maximizing cumulative reward for the agent over the curse of the problem. 22. What is the important task of reinforcement learning? The important task of reinforcement learning is to use rewards to learn a successful agent function.

4 PART - B 1. Explain about Learning From Observations. FORMS OF LEARIG: A learning agent can be thought of as containing a Performance element that decides what actions to take and a learning element that modifies the performance element so that it makes better decisions. The design of a learning element is affected by three major issues:- Which components of the performance element are to be learned? What feedback is available to learn these components? What reputation is used for the components? The components of the agents include the following:- A direct mapping from conditions on the current state to actions. A means to infer relevant properties of the world from the percept sequence. Information about the way the world evolves and about the results of possible actions the agent can take. Utility information indicating the desirability of world status. Action-value information indicating the desirability of actions. Goals that describe classes of states where achievement maximizing the agents utility. Each of the components can be learned from appropriate feedback. Consider, for example, an agent training to become a taxi driver. Every time the instructor shouts Break! the agent can learn a condition-action rule for when to breaks (component 1). By seeing many camera image that it is told contain human, it can learn to recognize them. By trying actions and observing the result-for example, breaking hard on a wet road-it can learn the effects of the action. Then, when it receives no tip from passengers who have been thoroughly shacks up during the trip, it can learn a useful component of it s overall utility function. The type of feedback available for learning is usually the most important factor is determining the nature of the learning problem that the agent faces.

5 The field of machine learning usually distinguishes three cases supervised, unsupervised and reinforcement learning. The problem of supervised learning involves learning a function from examples of it s inputs and outputs. Cases (1), (2) and (3) are all instances of supervised learning problems. In (1), the agent learns condition-action rule for breaking- this is a function from states to a Boolean output (to breaks or not to breaks). In (2), the agent learns a function from images to a Boolean output(whether the images contain a bus). In (3), the theory of breaking is a function from status and breaking action to say, stopping distance is feet. The problem of unsupervised learning involves learning pattern is the input when no specific output values are supplied. For example, a taxi agent might gradually develop a concept of good traffic days and bad traffic days without ever being given labeled examples of each. A purely unsupervised learning agent cannot learn what to do, because it has no information on to what constitutes a current action on a desirable state. The problem of reinforcement learning is the most general of three categories. Rather than being told what to do by a teacher, a reinforcement learning agent must learn from reinforcement. (Reinforcement learning typically includes the sub problem of learning how the environment works) The representation of the learned information also plays a way important role is determining how the learning algorithm must work The last major factor is the design of learning system is the availability of prior knowledge. The majority of learning reach is AI, computer science, and psychology has studied the case is which the agent begins with no knowledge at all about what it is trying to learn. Inductive learning: An algorithm for deterministic supervised learning is given as input) the current value of the unknown function for particular inputs and (must try to become the unknown function or some-thing close to it We say that an example is a pair (x,f(x)), where x is the input and f(x) is the output of the

6 function applied to x. The task of pure inductive inference (or induction) is this, Given a collection of examples of f, return a function h that approximates f. The function h is called a hypothesis. The reason that learning is difficult, from a conceptual point of view, is that it is not easy to tell whether any particular h is a good approximation of f (A good hypothesis will generalize well-that is, will predict example correctly. This is the fundamental problem induction) LEARIG FROM DECISIO TREES: Decision tree induction is one of the simplest and yet most successful forms of learning algorithm. Decision tree as performance elements:- A decision tree takes as input or object or situation described by a set of attributes and returns a decision,-the predicted output value for the input. The input attributes can be discrete or continuous. For now, a we assume discrete inputs. The output values can also be discrete or continuous; learning a discrete-valued function is called regression. We will concentrate a Boolean classification, where in each example is classified as true (positive) or false (negative). A decision tree reaches its decision by performing a sequence of tests. Each internal node is the tree compounds to a test of the values of one of the properties, and the branches from the node are labeled with the possible values of the cost. Each leaf node in the tree specifies the values to be returned if that leaf is reached.

7 Example:- The problem of whether to wait for a table at a restaurant. The aim here is to learn a definition for the goal predicate will wait. Patrons one Some Full o es Wait Estimate > o Alternate Hungry es Reservation? Fri/sat es Alternate? o es o es o es Bar? es o es o Raining? o es o es o Decision tree We will see how to automate this task, for now, let s suppose are decide on the following list of attributes: 1. Alternate: Whether there is a suitable alternative restaurant nearly. 2. Bar: Whether the restaurant has a comfortable bar area to wait is. 3. Fri/Sat: True on Fridays & Saturdays.

8 4. Hungry: Whether we are hungry. 5. Patrons: How many people are in the restaurant (Values are ame, Some &Full). 6. Price: The restaurants price range ($,$$,$$$). 7. Raining: Whether it is raining outside. 8. Reservation: Whether we made a reservation. 9. Type: The kind of restaurant (French, Italian, Thai or Burger). 10. WaitEstimate: The wait estimation by the host (0-10, 10-30, 30-60,>60). The decision tree usually used by one of us (SR) for this domain is given the Figure 4.1. The tree dose not we the price & Type attributes, is effect considering them to be irrelevant. Examples are processed by the tree starting at the root & following the appropriate branch until a leaf is reached. For instance, an example with patrons=full and Wait Estimate=0-10 will be classified positive. 2. Explain about Ensemble Learning. The idea of ensemble learning methods is to select a whole collection of hypothesis from the hypothesis space and combine these predictions. Methods are:- 1. Boosting 2. Bagging Motivation:- The motivation for ensemble learning is simple. Consider an ensemble (collection) of H= 5 hypothesis and suppose that their prediction are combined using simple majority voting. For the ensemble to misclassify it. The hope is that is much likely then a misclassification by a single hypothesis. Furthermore, suppose assume that the made by each hypothesis are independent. 1. In that case, if p is small then the probability of a misclassification occurring is very small. ow, obviously if different, thereby reducing the correlation between their ever, then ensemble learning can be very useful. 2. Ensemble idea is a generic way of enlarging the hypothesis space. That is, the ensemble itself can be thought of as a hypothesis and the new hypothesis space as the set of all possible ensemble construction from hypothesis is the original space.

9 If the original hypothesis space allows for a simple and efficient learning algorithm, then the ensemble method provides a way to learn a much more expressive class of hypothesis without much additional computational or algorithmic complexity. Boosting:- Boosting is the most widely used ensemble method. To understand how it work. It is necessary to understand the concept of a weighted training set. Weighted training set:- In a weighted training set, each example has an associated weight w j 0. The higher the weight of an example, the higher is the importance attached to it during the learning of a hypothesis. Working:- Boosting starts with w j =1 for all the examples. From this set, it generates the first hypothesis h 1. This hypothesis will classify some of the training examples correctly and some incorrectly. The hypothesis can be made better by increasing the weight of misclassified Examples while decreasing the weight of the correctly classified examples. From this new weighted training set, hypothesis h 2 is generated. The process continuous until M hypothesis has been generated, where M is an input to the boosting algorithm. The final ensemble hypothesis is a weighted-majority combination of all the AI hypothesis, each weight according to how well it performed on the training set. ADABOOST algorithm:- ADABOOST algorithm is one among the variants. ADABOOST is one of the most commonly used boosting algorithms. It has a important property is which if the input learning algorithm L is a weak learning algorithm, which means that L always returns a hypothesis with weighted ever on the training set.

10 LOGICAL FORMULATIO OF LEARIG Expressions of Decision tree:- Any Particular decision tree hypothesis for the Will wait goal predicate can be seen as an aeration of the form, s will wait(s) (p 1 (s) p 2 (s) p n (s)), Where each condition p i (s) is a conjunction of tests corresponding to a path from the root of the tree to a leaf with a parities octane. The decision tree is really deciding a relationship between Will Wait and some logical combination of attribute values. We cannot decision tree to represent tests that refer to two or more different objects-for example, Э r2 ear by (r 2,r) Price(r, p) Price(r 2,p 2 ) Cheaper(p 2,p) Decision tree are fully expressive within the class of proportional language, that is, any Boolean function can be written as a decision tree. This can be done trivially by having each row is the truth table for the function corresponds to a path in the tree. This would yield an exponentially large decision tree representation because the truth table has exponentially many rows. Clearly, decision tree can represent many functions with much smaller tree. Clearly, decision tree can represent much function with much smaller tree. For some kinds of functions, however this is a real problem. For example, if the function is the parity function, which returns 1 if and only if an every number of inputs are 1, then an exponentially large decision tree will be needed. It is also difficult to represent a majority function, which returns 1 if more than half of its inputs are 1. The truth table has 2 n rows, because input care is described by n attributes. We can consider the answer column of the table as a 2 n -bit number that defines the function. If it takes 2 n bits to define the function, then there are 2 2ˆn different functions on n attributes. For an, with just six Boolean attributes, there are 2 2ˆ6 =18,446,744,673,709,551,616 different function to choose from.

11 Inducing decision tree from example:- An example for a Boolean decision tree consists of a vector of input attributes, X and a single Boolean output value y. A set of examples (x 1,y 1 ),(x 12,y 12 ) is shown in the Figure Example Attributes Goal X 1 Alt Bar Fri Hun Pat Price Rain Re s Some $$$ Type Est Willwait French 0-10 X 2 X 3 Full $ Thai 30- X 4 60 X 5 Some $ Buya X X 7 Full $ Thai X X 9 Full $$$ French 30 X 10 X 11 Some $$ Italian >60 X 12 one $ Buya 0-10 Some $$ Thai 0-10 Full $ Buya 0-10 Full $$$ Italian >60 one $ Thai 10-30

12 Full $ Buya 0-10 Figure 4.2 Inducing decision tree from example The positive example are the once in which the goal Will Wait is true (x 1, x 3, ); the negative examples are the ones is which it is false (x 2,x 5, ). The complete set of example is called training set. Trivial tree:- The problem of finding a decision tree that agree with tracing set might seem difficult, but is fact there is a trivial solution that is constructing trivial tree. Construct a decision tree that has one path to a leaf for each example, where the path tests each attributes is turn and follows the classification of the example. Problem with trivial tree:- It just memorizes the observations. It does not extract any pattern from the example, so it is not expected to be able to extrapolate to example it has not seen. The above Figure 4.2 shows how the algorithm gets started 12 training example are given, which are classified into positive and negative sets. Then which attribute to be on the first test is the tree is decided, because it leaves in with four possible outcomes, each of which has the same number of +-ve & -ve examples. On the other hand, the following Figure 4.2. Shown that patrons is a fairly important attribute, There are four cases to consider for the recursive problem:- 1) If there is some positive & negative example, then choose the best attribute to split them. Figure 4.2 (h) shows Hungry being used to split the remaining examples. 2) If all the remaining example are positive, then it is done: can answer es or o, Figure(a) shows examples of this is the ame & Some cases. 3) If there are no examples left, it means that no such example has been observed, and a default values calculated from the majority classification at the nodes parent is returned.

13 4) If there are no attributes left, but both positive and negative example, then there examples have exactly the same description, but different classification. The decision-tree-learning algorithm:- Function DECISIO-TREE-LEARIG(example, at ribs, default)returns a decision tree Input: example, attains, default If example is empty then return default Else if all examples have the same classifications then return the classification else if attributes is empty then return MAJORIT-VALUE (example) Else Best CHOOSSE-ATTRIBUTE (attibs, examples) Tree a new decision tree with cost test best M MAJORIT-VALUE(EXAMPLES) For each values of V i best do Example {element of example with best=v i } Subtree DECISIO-TREE-LEARIG(example, attribs-best, m) Add a branch to tree with label u, and sub tree Return tree The final tree produced by the algorithm applied to the 12-example data set is shown in Figure 4.2. Choosing attribute test:- To choose attribute test, the idea is to pick the attribute that goes as for as possible toward providing an exact classification of the examples. A perfect attribute divides the examples into sets that are all positive or all negative. Measure to find finely good and really useless attributes:- In general, if possible ensure V i have probabilities P(V i ), then the information content I of the actual answer is given by, I(p(v i ), p(v n ))= p(v i )log 2 p(v i ) To check this equ, for the tossing of a fair coin, the following can be used, I[1/2,1/2]=-1/2log1/2-1/2log 2 1/2=1bit

14 Gain(A)=I[p/p+n,n/p+n]-Remainder(A) Gain(patterns)=1-[(2/12)I(0,1)+(4/12)I(0,1)+(6/12)I(2/6,4/6)] ~0.54bits Gain(Type)=1-[(2/12)I(1/2,1/2)+(2/12)I(1/2,1/2)+(4/12)I(2/4,2/4)+(4/12)I(2/4,2/4)] =0 Assume the performance of the learning algorithm:- A learning algorithm is good if it produces hypothesis that do a good job of predicting the classification of example prediction quality can be estimated in advance on it can be estimated after the fact. 1) Collect a large set of example. 2) Divide it into two disjoint sets, the training set and the test set. 3) Apply the learning algorithm to training set, generating a hypothesis h. 4) Measure the percentage of example in the test set that is correctly classified by h. 5) Repeat step 1 to 4 for different size of training sets and different randomly selected training sets of each size. oise and over fitting:- The problem of finding meaningless regularity in the data, whenever there is a large set of possible hypothesis is called over fitting. Solution:- Decision tree pruning Cross validation 3. Explain about KOWLEDGE-I-LEARIG:- Prior knowledge can help an agent to learn from new experiences. Logical formulation of learning:- Inductive learning, which is a process of finding a hypothesis that, agrees with the observed example Hypothesis is represented by a set of logical sentences Logical sentences such as prior knowledge, example description and classification Thus, logical inference aids learning Examples and hypothesis:-

15 Restaurant learning problem i.e., learning a rule for deciding whether to wait for a table. The example object is described by logical sentence, the attribute is way predicates.

16 Example alternate Bar Fri Hungry Patern Price Rain Rearch Est Goal will wait X 1 Some $$ o es 0-10 yes Figure 4.3 Knowledge-in-learning Di(X i )=Alternate(X 1 ) Bar(X 1 ) Fri/Sat(X 1 ) Hungry(X 1 ) Patern(X 1 ) Price($$$) Rain(X 1 ) Research(X 1 ) Where D i logical expression taking single argument. Classification of the object is done by WillWait(X 1 ) on the generic notation is, Q(X 1 ) if the example is positive. Q(X 1 ) if the example is negative. The complete training set is just the conjunction of all description and classification sentences. The hypothesis can propose an expression called candidate definition of the goal predicate. C i =candidate definition. Then hypothesis H i is a sentence of the form, x Q(x) C i (x) For a decision tree, the goal predicate of an object is leading to tree is satisfied. The decision tree induced from the training set, r willwait (r) Pations (r, Some) Pations (r, full) Hungry(r) Type (r, French) Pations (r, full) Hungry(r) Type(r, thai) Fri/Sat(r) Pations (r, full) Hungry(r) Type(r, Buyer) If is a function of branches where output on leaf node is true.

17 Current best hypothesis search:- To maintain single hypothesis. It is adjusted to maintain the consistory. This search algorithm is painted by John Stuart Mill(1843) Generalization:- Hypothesis says it should be negative but it is true. Hence this entering has to be included in the hypothesis. This is called Generalization. Specialization:- Hypothesis says that the new example is positive, but it is actually negative. This criterion must be decreased to exclude this example. This is called Specialization. Current-best learning algorithm is used in many machine-learning alg. 4. Explain about Explanation Based Learning (EBL):- Definition:- When an agent can utilize a worked example of a problem as a problem-solving method, the agent is said to have the capability of Explanation-based learning (EBL). Advantage:- A deductive mechanism, it requires only a single training example. EBL algorithm requires all of the following:- Accepts 4 kinds of inputs:- 1. A training example:- What are the learning seen in the world. 2. A goal concept A high level description of what the program is to learn. An operational criterion:- A description of which concept are usable. A domain theory:- A set of rule that describe relationship b/w object and action is a domain. Entailment constraint satisfied by EBL is, Hypothesis description classifications

18 Background hypothesis Extracting rules from example:- If is a method for extracting general rules from individual observations. Idea is to construct an explanation of the observation using prior knowledge. Consider (X 2,X)=2X Suppose to simply 1X(0+X). the knowledge here include the following rules, Rewrite (u,v) Simplify (v,w)=>simplify(u,w) Primitive(u)=>Simplify(u,u) Arithmetic unknoun=>primitive(u) umber(u) primitive(u) Rewrite(1*u,u) Rewrite(0+u,u) EBL process working:- 1. Construct a proof that the goal predicted applies to the example using the available background knowledge. 2. In parallel, construct a generalized proof tree for the variabilized goal using the same inference steps as in the original proof. 3. Construct a new rule where left hand side consists of leaves of the proof tree and RHS is the variabilized goal. 4. Drop any conditions that are tree regardless the values of the variable is the goal. Improving efficiency:- 1. Reduce the large number of rules. It increases the branching factor in the search space. 2. Derived rules must offer significant increase in speed. 3. Derived rule is a general as possible, so that they apply to the largest possible set of cases.

19 5. Explain about Learning Using Relevance IFORMATIOS:- Relevance based learning using prior knowledge. Eg:- traveler is Brazil concluded that all Brazilian speaks Portuguese. Entailment constraint is given by, Hypothesis Description Classification Background Description Classification Hypothesis Express the above example is FOL as, ationality (x,n) ationality (y,n) language (x,l) => Language (y,l) If x & y have the same nationality & x speaks language l & then y also speaks it. ationality (Fernando, Brazil) language (Fernando, Portuguese Entails the following sentence as, ationality (x, Brazil) => Language (x, Portuguese) The relevance for given nationality language is fully determined. There sentences are called functional dependencies or determination. ationality (x, n) => language (x, l) IDUCTIVE LOGIC PROGRAMMIG:- Knowledge based inductive learning (KBIL) find inductive hypothesis that explain set of observation with the help of background knowledge. ILP techniques perform KBIL on knowledge that expressed in first order logic. ILP gained popularity for three reasons:- 1. It offers rigorous approach to general knowledge-based inductive learning problem. 2. It offers complete algorithm for including general first order theories from examples. 3. It produces hypothesis that are easy for humans to read. Ex:- Problem for learning family relationship. The descriptions consists of mother, father, married relation, male and female properties. A typical family tree:- The corresponding descriptions are as follows, Father (Philips, Charles) Father (Philips, Anne). Mother

20 (Mum, Margaret) Mother (Mum, Elizabeth). Married (Diana, Charles) Married (Elizabeth, Philip) Male (Philip) Male (Charles). Female (Diana) Female (Elizabeth). Grandparent (Mum, Charles) Grandparent (Elizabeth, Bectria). Grandparent (Mum, Harry) Grandparent (Spencer, Peter) Statistical learning methods:- The key concept of statistical learning is domain & hypothesis. (Domain work) Data:- Data are evidence of some or all of the random variables describing the domain.

21 Hypothesis:- The hypothesis is probabilistic theories of how the domain works including logical theories as a special case. Example; Candy comes as two flavors: cherry & lime Bayesian learning: It calculates the probability of each hypothesis, given the data & makes predication by use all the hypothesis. Bayesion view of learning is extremely powerful, providing general solution the problem of noise, over fitting & optional prediction. Ex:- h 1 :100% cherry h 2 :75% cherry +25% lime h 3 :50% cherry + 50% lime h 4 :25% cherry +75% lime h 5 :100% lime Learning with complete data:- Parameter learning task involves finding the numerical parameter for the probability model. The structure of the model is fixed. Data are complete when each data point contains values for every variable in the probability model. Complete data simplify the problem of learning the parameter of complex model. REIFORCEMET LEARIG:- If involves finding a balance b/w exploration of new knowledge and exploitation of current knowledge. Ex:- chess by supervised learning.

22 o agent means, it tries some random moves. The agent need to know that something good has happened when it wise and something bad when it loss. This kind of feedback is called leeward or reinforcement. In game like chess, the reinforcement is received only at the end of the game. In ping-pong game each point scored can be considered as leeward. The test of reinforcement learning is to be observed rewards to learn an optional poling for the environment. It is a sub-area of machine learning. The basic reinforcement learning model consists of:- 1. A set of environment states s; 2. A set of action A; and 3. A set of scalar rewards ; R If is well suited for application like robot control, telecommunication and games. Various types of agents:- 1. Utility based agents:- It learns utility function on states, and cases it to select action that maximizes the expected outcome utility. 2. Q-learning agent:- It learns an action-value function or Q- function, giving the expected utility of taking a given action is a given state. 3. Reflex agent:- It learns a policy that maps directly from states to action. Basic reinforcement learning model: Types of reinforcement learning: