Knowledge Representation and Reasoning Course Outline PDF

Knowledge Representation and Reasoning Rafael Peñaloza Milano, 2024-2025 Università di Milano-Bicocca The Course Schedule: Mondays 8:30 – 11:30 Thursdays 11:30 – 14:30 Room U1-09 Moodle KIRO https://elearning.unipv.it/ All info...

Knowledge Representation and Reasoning Rafael Peñaloza Milano, 2024-2025 Università di Milano-Bicocca The Course Schedule: Mondays 8:30 – 11:30 Thursdays 11:30 – 14:30 Room U1-09 Moodle KIRO https://elearning.unipv.it/ All information, calendar, and material available there KRR 2024-2025 2 Programme What is KR and where does it come from? Boolean algebra / Basics of Logic KR as rules; basic logic programming Existential rules and Kripke semantics of FOL EL: a simple DL The DL ALC Other DL constructors; trade-offs Linear temporal logic; Satisfiability of LTL formulas Uncertainty; ProbLog KRR 2024-2025 3 Distribution approx. 32 hours lecture + 24 hours exercises Consult the calendar KRR 2024-2025 4 Exam Written Exam at the end of the semester If you can solve the exercises, you should be ok Final grade combined with Module 2 more information towards the end of the semester KRR 2024-2025 5 Lecture Notes No reference book (at the moment) Material, references, and exercises will be given throughout the semester Rewards if you spot errors KRR 2024-2025 6 Scope of the course Understand KR and symbolic AI (of course) Get tools to critically evaluate methods, results, and claims Introduce mathematical thinking and understanding KRR 2024-2025 7 Doubts? Ask questions, ask early, ask through the elearning forum Office Hours: Mondays 14:00–15:00 (under appointment) KRR 2024-2025 8 Suggestions The course is difficult and needs to be digested Follow the material on time Ask questions Try the exercises by yourselves Do other things! KRR 2024-2025 9 Logic and Symbolic AI Artificial Intelligence What is intelligence? Artificial Intelligence as a whole deals with machines showcasing intelligent traits This is not necessarily equivalent to (but often intersects with) simulating human cognitive attributes KRR 2024-2025 10 Thinking Fast and Slow Two separate thought processes System 1 System 2 fast slow automatic effortful subconscious conscious “intuitive” “logical” reflexes purpose KRR 2024-2025 11 Hardwired System 1 is a “hard-wired” mechanism which immediately reacts to stimuli Very useful for simple tasks and pattern following but can be easily fooled Creating it and modifying it can be very expensive rigid structure KRR 2024-2025 12 Pareidolia KRR 2024-2025 13 The Current Hype We all get the news AI system for: writing text generating images winning games detecting diseases optimising results classifying images... Based on System 1 Machine Learning / sub-symbolic AI KRR 2024-2025 14 Drawbacks of Deep Learning Deep learning is very useful for many things but we should be aware of its limitations Some obvious drawbacks: lack of interpretability “hallucinations” (always produces an answer) static (non-updateable) disregard existing knowledge A deep learning “model” is just an opinion stochastically formed on the data it digested KRR 2024-2025 15 No double check No process to refine verify correct the perception or generation KRR 2024-2025 16 Symbolic AI Symbolic AI is the analogous of System 2 It manipulates symbols to extract consequences and make decisions In other words: it performs reasoning (valid inferences) what can I deduce given the knowledge? Mainly interested in guarantees only answers which are certain KRR 2024-2025 17 Advantages interpretability correctness guarantees flexibility modularity “humbleness” no ambiguity KRR 2024-2025 18 This course is about symbolic AI KRR 2024-2025 19 Logic You will hear me speak a lot about logic Do not fret! a “logic” is just a “language” without the nuances we use it to express and manipulate knowledge It associates expressions with their meaning KRR 2024-2025 20 Reasoning Remember: reasoning is all about guarantees we must be certain of the answer! Important for critical applications KRR 2024-2025 21 Example Ana, Bob and Claudia want to organise a party: Ana wants Daria or Eric (or both) to be present; Bob doesn’t like Eric and Fabiola; at least one of them should be excluded Claudia does not want Daria, unless Fabiola is present Is it possible to organise the party? Who is invited? What else can we conclude? KRR 2024-2025 22 Reasoning This is a typical example of a reasoning problem The constraints of all organisers must be satisfied (no approximations or heuristics) Answering it requires some careful analysis KRR 2024-2025 23  Example - Solution 1. Ana: Daria or Eric (or both) present; 2. Bob: Eric or Fabiola (or both) excluded; 3. Claudia: if Daria, then also Fabiola is present Invite Daria Invite Eric then Fabiola must come (3) then Fabiola cannot come (2) then Eric is not invited (2) then Daria is not invited (3) Daria and Fabiola Eric only Is George invited? KRR 2024-2025 24 The Basics The “party” problem belongs to a class of resource allocation or constraint satisfaction problems Can be modelled (and solved) in the abstract through the very simple propositional logic Although the constraints can get very complicated When we reason, we do symbolic manipulation the meaning of the symbols is irrelevant KRR 2024-2025 25 Expressivity While useful, propositional logic cannot represent all kinds of knowledge Gives rise to many different representation languages different properties require different treatment We cannot just make one big “unified” language for practical reasons Impossible to study them all consider prototypical examples and extensions KRR 2024-2025 26 Goal Give you the tools to understand when to use KR, use it effectively, and avoid unnecessary overhead KRR 2024-2025 27 [email protected]