Artificial Intelligence

Ruth Aylett
Lilia Georgieva
 Our objectives
 Critical understanding of traditional AI problem solving and knowledge
representation methods
 Use of knowledge representation techniques
 Critical understanding of different systematic and heuristic search
techniques
 Practice in expressing problems in terms of state-space search
 Broad knowledge and understanding of the subfields and applications
of AI e.g computer vision, machine learning and expert systems.
 Detailed knowledge of one subfield of AI (e.g. natural language
processing, planning) and ability to apply its formalisms and
representations to small problems
 Detailed understanding of different approaches to autonomous agent
and robot architectures, and the ability to critically evaluate their
advantages and disadvantages in different contexts.
 Practice in the implementation of simple AI systems using a suitable
language
 The course - 1
 Lilia will cover:
– Search algorithms (depth first search, breadth first
search, uniform cost search, A* search), including
complexity and implementation;
– constraint satisfaction problems;
– games (min-max, alpha-beta pruning);
– logic, resolution, introductory logic programming
– reasoning with uncertainty (Bayesian networks,
probability theory)
– computer vision
Materials at www.macs.hw.ac.uk/~lilia/AI
 The course - 2
 Ruth will cover:
–rule-based systems and programming
–planning
–agents (software, graphical, affective)
–machine learning (rule induction, GAs)
–non-symbolic agents (biologically-inspired)
–knowledge-based systems (expert systems),
knowledge engineering
Materials via www.macs.hw.ac.uk/~ruth/teaching.html
 Books
 Artificial Intelligence - a modern approach
• Russell & Norvig (Prentice Hall)
• http://aima.cs.berkeley.edu/
– Originally used Lisp but some code now available
in Java and Python
 Artificial Intelligence: Structures and
Strategies for Complex Problem Solving
• George Luger (Pearson (Addison-Wesley))
• http://www.cs.unm.edu/~luger/ai-final/
– Prolog, Lisp and some Java code available
 Essence of Artificial Intelligence
• Alison Cawsey - of Heriot-Watt
• Very clear explanations
 Assessment
 Two pieces of in-course assessment
(30%)
– From Lilia: will be handed in mid-semester
with demos/feedback in week 7
– From Ruth: will be handed in end-of
semester with demos/feedback in week 12
 Strongly related to your lab-work!
 One written exam (70%)
 Labs
 Lilia
– Pandora tool
– SPASS theorem prover
 Ruth
– JESS (Java Expert System Shell)
– JADE (Java Agent Development
Environment)
 Early Artificial Intelligence
 Alan Turing, early 1950s
– The Turing Test
 Simon and Newell, late 1950s
– General Problem Solver - GPS
 Shakey - Stanford 1966-72
– First autonomous robot
– STRIPS planner
 Different approaches
 No agreement on what we are doing..
 Philosophical
– Definition of ‘intelligence’
• To recognise alien intelligence maybe?
 Cognitive
– Plausible models of human abilities
• To better understand humans
 Engineering
– Behaviour which is ‘intelligent’ for humans
• Chess? Recognising a pencil?
• Why humans?
 Be able to Do the Right Thing!
 Changing aims...
From: Powerful, general mechanisms
– ‘General’ Problem Solver (1950s-60s)
Via: Knowledge-rich, domain specific
mechanisms
– Expert Systems; Cognitive modelling
(1970s-80s)
To: Situated intelligence
– Integrating sensing and acting (1990s-00s)
 Different Aims
 Create an Artificial Intelligence?
– As good as (or better than!) a human?
– STRONG AI
– Argued that this is definitionally impossible
(and also possibly immoral or dangerous)
 Create better artefacts
– In limited domains
– Certainly possible but not very exciting
 Strong AI and the Chinese
Room
 A ’thought’ experiment
– John Searle, a philosopher
 A room, a person with some rules for
translating inputs to outputs
– Inputs are some Chinese characters
– Rules used to produce output Chinese characters
 But does anyone ‘understand’ Chinese?
http://www.iep.utm.edu/c/chineser.htm
 Good Old Fashioned AI
(GOFAI)
Three major components:
1. Explicit representation of knowledge
– Usually symbolic
– Supports explanation
2. Inferencing Mechanisms
– Application of knowledge
– Production of new knowledge
3. Current conclusions
– Working memory
 Types of knowledge - 1
 Declarative
– Objects and relationships
– General, static
– Maps, Diagrams, Definitions, Anatomy
 Procedural
– Processes, instructions
– Specific, dynamic
– Recipes, DIY instructions, Surgery
 Types of Knowledge - 2
 Heuristic
– Partial, ‘rule of thumb’, pragmatic
– ‘Surface’ knowledge
 Model-based
– Structure, relationships
– ‘Deep’ knowledge
 A standard approach
 Represent the domain knowledge
– With an appropriate symbolic
representation
 Search the space of solutions
– Using an appropriate heuristic method
– Until an ‘acceptable’ solution is found
 An example
 The fox, the goose and the cabbage
A River B

 Aim: to get all three from A to B

– Constraint 1: only room for one at a time
– Constraint 2: cannot leave fox with goose or
goose with cabbage
 Representing the problem

a f, c
 Describing a state boat a g

 Operators:
– LB ?object (Load Boat)
– CR ?from ?to(Cross River)
– ULB ?object (Unload Boat)
 Searching the space
 Iterative process start
– Which operators can be
applied in given state? LB,f LB,g LB,c CR

• Validity; avoid looping S1 S2 S3 S4

– What state results from each

operator?
– Are any of these the answer?
 Finding the solution
 For any branch terminating in goal state
– Sequence of operators in that branch is
solution
 Solution characteristics
– One or many?
– Any or best?
• ‘best’ = fewest operators?
• Or ‘cheapest’ operators?
 Inferencing style
 Goal-driven reasoning
– From desired end-state backwards
– Start with a hypothesis
– Or assume the answer is true…
 Data-driven reasoning
– From start-state forwards
 Did it matter in this problem?