1	Foundations of Artificial Intelligence Revision
2	Final Exam Venue: PGP General Purpose Room Date: 23 April (Friday) Time: 2:30 – 4:30 pm
3	Format One A4 sized sheet allowed to the test Eight questions, emphasizing material covered after the midterm Yes, all material in the course will be covered on the exam
4	No class next week Today is the final lecture for the course You had your “extra” lecture in webcast as the vision, NLP or robotics advanced topics lecture
5	Outline Agents Search Uninformed Search Informed Search Adversarial Search Constraint Satisfaction Knowledge-Based Agents Uncertainty and Learning
6	Agent types Four basic types in order of increasing generality: Simple reflex agents Model-based reflex agents Goal-based agents Utility-based agents
7	Simple reflex agents
8	Model-based reflex agents
9	Goal-based agents
10	Utility-based agents
11	Creating agents Where does the intelligence come from? Coded by the designers Knowledge representation – predicate and first order logic Learned by the machine Machine learning – expose naïve agent to examples to learn useful actions
12	Learning agents
13	Searching for solutions In most agent architectures, deciding what action to take involves considering alternatives Searching is judged on optimality, completeness and complexity Do I have a way of gauging how close I am to a goal? No: Yes:
14	Uninformed search Formulate the problem, search and then execute actions Apply Tree-Search For environments that are Deterministic Fully observable Static
15	Tree search algorithm Basic idea: offline, simulated exploration of state space by generating successors of already-explored states
16	Summary of algorithms Breadth-First – FIFO order Uniform-Cost – in order of cost Depth-First – LIFO order Depth-Limited – DFS to a maximum depth Iterative Deepening – Iterative DLS. Bidirectional – also search from goal towards origin
17	Repeated states: Graph-Search
18	Informed search Heuristic function h(n) = cost of the cheapest path from n to goal. Greedy Best First Search Minimizing estimated cost to goal A* Search Minimizing total cost
19	Properties of heuristic functions Admissible: never overestimates cost Consistent: estimated cost from node n+1 is than cost from node n + step cost. A* using Tree-Search is optimal if the heuristic used is Graph-Search needs an consistent heuristic. Why?
20	Local search Good for solutions where the path to the solution doesn’t matter Often work on Don’t search systematically Often require very little memory Correlated to online search Have only access to the local state
21	Local search algorithms Hill climbing search – choose best successor Beam search – take the best k successor Simulated annealing – allow backward moves during beginning steps Genetic algorithm – breed k successors using crossover and mutation
22	Searching in specialized scenarios Properties of the problem often allow us to formulate Better heuristics Better search strategy and pruning Adversarial search Working against an opponent Constraint satisfaction problem Assigning values to variables Path to solution doesn’t matter View this as an incremental search
23	Adversarial Search Turn-taking, two-player, zero-sum games Minimax algorithm: One ply: Max nodes: agent’s move, maximize utility Min nodes: opponent’s move, minimize utility Alpha-Beta pruning: rid unnecessary computation.
24	Constraint Satisfaction Discrete or continuous solutions Discretize and limit possible values Modeled as a constraint graph As the path to the solution doesn’t matter, local search can be very useful.
25	Techniques in CSPs Basic: backtracking search DFS for CSP A leaf node (at depth v) is a solution Speed ups Choosing variables Minimum remaining values Most constrained variable / degree Choosing values Least constraining value
26	Pruning CSP search space Before expanding node, can prune the search space Forward checking Pruning values from remaining variables Arc consistency Propagating stronger levels of consistency E.g., AC-3 (applicable before searching and during search) Balancing arc consistency with actual searching.
27	Propositional and First Order Logic Propositional Logic Facts are true or false First Order Logic Relationships and properties of objects More expressive and succinct Quantifiers, functions Equality operator Can convert back to prop logic to do
28	Inference in logic Given a KB, what can be inferred? Query- or goal-driven Backward chaining, model checking (e.g. DPLL), resolution Deducing new facts Forward chaining Efficiency: track # of literals of premise using a count or Rete networks
29	Inference in logic Chaining Requires Uses Modus Ponens for sound reasoning Forward or Backward types Resolution Requires Conjunctive Normal Form Uses Resolution for sound reasoning Proof by Contradiction
30	Inference in FOL Don’t have to propositionalize Could lead to infinite sentences functions Use unification instead Standardizing apart Dropping quantifiers Skolem constants and functions Inference is semidecidable Can say yes to entailed sentences, but non-entailed sentences will never terminate
31	Connection to knowledge-based agents CSP can be formulated as logic problems and vice versa CSP search as model checking Local search: WalkSAT with min-conflict heuristic
32	Inference and CSPs Solving a CSP via inference Handles special constraints (e.g., AllDiff) Can learn new constraints not expressed by KB designer Solving inference via CSP Whether a query is true under all possible constraints (satisfiable) Melding the two: Constraint Logic Programming (CLP)
33	Uncertainty Leads us to use probabilistic agents Only one of many possible methods! Modeled in terms of random variables Again, we examined only the discrete case Answer questions based on full joint distribution
34	Inference by enumeration Interested in the posterior joint distribution of query variables given specific values for evidence variables Summing over hidden variables Cons: Exponential complexity Look for absolute and conditional independence to reduce complexity
35	Bayesian networks One way to model dependencies Variable’s probability only depends on its parents Use product rule and conditional dependence to calculate joint probabilities Easiest to structure causally From root causes forward Leads to easier modeling and lower complexity
36	Learning Inductive learning - based on past examples Learn a function h() that approximates real function f(x) on examples x Balance complexity of hypothesis with fidelity to the examples Minimize α E(h,D) + (1-α) C(h)
37	Learning Algorithms Many out there but the basics are: K nearest neighbors Instance-based Ignores global information Naïve Bayes Strong independence assumption Scales well due to assumptions Needs normalization when dealing with unseen feature values Decision Trees Easy to understand its hypothesis Decides feature based on information gain
38	Training and testing Judge induced h()’s quality by using a test set Training and test set must be separate; otherwise peeking occurs Modeling noise or specifics of the training data can lead to overfitting Use pruning to remove parts of the hypothesis that aren’t justifiable
39	Where to go from here? Just the tip of the iceberg Many advanced topics Introduced only a few Textbook can help in exploration of AI
40	That’s it Thanks for your attention over the semester See you in April!