Intelligence and Knowledge

University of Ulster Faculty of Informatics Postgraduate Diploma in Educational TechnologyEducational and Management Support System

Intelligence and KnowledgeINTELLIGENCE AND KNOWLEDGE.................................................................................................................... 1

1. RECOGNISING INTELLIGENCE.............................................................................................................................. 21.1 Turing test..................................................................................................................................................... 21.2 Loebner Competition.................................................................................................................................... 21.3 ThoughtTreasure............................................................................................................................................ 31.4 Newell and Simon’s definition of intelligence................................................................................................31.5 Physical Symbol System hypothesis............................................................................................................... 31.6 Church-Turing thesis..................................................................................................................................... 4

2 MACHINE LEARNING........................................................................................................................................... 43 KNOWLEDGE BASED SYSTEMS............................................................................................................................. 4

3.1 Components of a KBS.................................................................................................................................... 43.2 How does the inference engine work? ( for production systems)....................................................................53.3 Control of search in production systems........................................................................................................ 53.4 When to go for an expert system solution......................................................................................................63.5 Other ways of representing knowledge.......................................................................................................... 73.6 Semantic networks........................................................................................................................................ 83.7 Frames and scripts....................................................................................................................................... 93.8 Assessing an expert system.......................................................................................................................... 103.9 Tools available to build expert systems.......................................................................................................10

4 THE KNOWLEDGE ENGINEERING PROCESS............................................................................................................ 104.1 Concerns of the expert................................................................................................................................. 114.2 Knowledge Acquisition............................................................................................................................... 114.3 What is knowledge?.................................................................................................................................... 124.4 Knowledge and expertise............................................................................................................................ 124.5 Types of knowledge..................................................................................................................................... 13Procedural knowledge ( knowing how)............................................................................................................. 13Declarative knowledge ( knowing that)............................................................................................................ 13Semantic knowledge.......................................................................................................................................... 13Episodic knowledge........................................................................................................................................... 134.6 Techniques for knowledge elicitation...........................................................................................................13Interviewing...................................................................................................................................................... 13Protocol analysis............................................................................................................................................... 13Multi-dimensional techniques............................................................................................................................ 14Goal decomposition techniques.......................................................................................................................... 14

5. ROBOCUP POSES A NEW SET OF AI RESEARCH CHALLENGES..........................................................................145.1 The RoboCup challenge............................................................................................................................... 145.2 Three different leagues............................................................................................................................... 145.3 Limited results in 1997................................................................................................................................. 155.4 Rules of the game......................................................................................................................................... 155.5 Middle-Size League..................................................................................................................................... 155.6 Small League............................................................................................................................................... 155.7 Freed from physical bodies.......................................................................................................................... 165.8 Simulation League....................................................................................................................................... 165.9 Getting better............................................................................................................................................... 175.10 The Future................................................................................................................................................. 17

KNOWLEDGE ELICITATION EXERCISE...................................................................................................................... 18

Intelligence and Knowledge

1 of 19


1. Recognising Intelligence

In 1950, Alan Turing said :

" I believe that in about 50 years' time it will be possible to program computers to make them play the imitation game so well that an average interrogator will not have more than a 70% chance of making the right identification after five minutes of questioning. The original question, 'Can machines think?' I believe to be too meaningless to deserve discussion. Nevertheless I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted"

1.1 Turing test

The test, which Turing call the imitation game, is aimed at measuring the performance of an allegedly intelligent machine against that of a human . The machine and a human counterpart are placed in rooms apart from a second human ( the interrogator). The interrogator's only means of communication is via a textual device such as a terminal and has to distinguish the computer from the human on the basis of the answers to questions asked via this device. If the interrogator cannot distinguish the machine from the human then the machine passes the test and may be assumed to be intelligent.

Important features of the Turing test It gives us an objective notion of intelligence. It prevents us being sidetracked by considerations such as whether the machine is conscious of its actions. It eliminates any bias in favour of living organisms.

Criticisms of the Turing test It does not test perceptual skills or manual dexterity.

This raises the question as to whether the Turing test constrains machine intelligence to fit human ideas of intelligence. The amount of knowledge that a machine would need to possess in order to pass the Turing test is huge . Some believe that it is impossible for a machine to pass the test.

1.2 Loebner Competition

The Loebner Prize Medal and a cash award is awarded annually to the designer of the computer system that best succeeds in passing a variant of the Turing Test. In 2000, $2,000 and a bronze medal will be awarded to the designer of the Most Human Computer as rated by a panel of judges.

The Loebner Prize Competition in Artificial Intelligence was established in 1990 by Hugh Loebner and the Cambridge (Massachusetts) Center for Behavioral Studies. To obtain transcripts and results of previous competitions see Loebner Prize Homepage. (www.loebner.net/Prizef/loebner-prize.html )

In accordance with the requirements of Dr. Loebner as published in the June 1994 Communications of the ACM, the winner of the $100,000 Grand Prize must be prepared to deal with audio visual input, and appropriate competitions will be held once competitors have reached Turing's 30:70 likelihood level of being mistaken for a human. An intermediate Turing Prize of $25000 will be offered for the first/best program to exceed this level.

You can talk to some of the winning programs at < http://cogsci.ucsd.edu/~asaygin/tt/ttest.html#talktothem >

Clearly it is difficult for a machine to pass the unrestricted Turing test. Perhaps this should not be too worrying since very few humans would be capable of carrying on a conversation on every topic imaginable! Provided the domain of discourse is restricted to being very narrow then there are many programs which can match the performance of humans, normally referred to as expert or knowledge based systems.

Of course some people are of the opinion

' there can be no machine intelligence without machine learning'

2 of 19

http://cogsci.ucsd.edu/~asaygin/tt/ttest.html#talktothem

http://www.loebner.net/Prizef/loebner-prize.html


and' human intelligence is approximately 99% pattern recognition and only 1% reasoning'.

1.3 ThoughtTreasure

Since 1994 E T Mueller has been developing ThoughtTreasure, a comprehensive platform for natural language processing (English and French) and commonsense reasoning. It runs on PCs and Unix and is available for free download. http://www.signiform.com/tt/htm/overview.htm ThoughtTreasure allows an application to obtain answers to questions easily answered by humans but previously difficult for computers such as:

Q: Do people have fingernails?. Q: What colour is the sky?. Q: How long does a play typically last? Q: How many husbands does a person have? Q: Can beds speak? Q: What do you do at the end of a phone call? Q: What do you call someone who is 16 years old? Q: What are rough synonyms for food?

Human commonsense knowledge is huge, probably consisting of over 100 million items. So far, ThoughtTreasure only contains 100 thousand items. But even a small amount of common sense will improve applications. For example, a calendar and contact manager can use ThoughtTreasure to: guess how long an appointment will last, understand that my mum and my mother refer to the same person, refer to the user's friends by first name and others by full name, and map various occupations and hobby descriptions to a shorter, standard list.

ThoughtTreasure contains a database of 20,000 concepts organized into a hierarchy. For example, Evian is a type of flat-water, which is a type of drinking-water, which is a type of beverage, which is a type of food, and so on.

Each concept has one or more English and French approximate synonyms, for a total of 50,000 words and phrases. For example, associated with the food concept are the English words food and foodstuffs and the French words aliment and nourriture (and others).

ThoughtTreasure contains 14,000 assertions about concepts, such as: a green-pea is green, a green-pea is part of a pod-of-peas, and a pod-of-peas is found in a typical grocery store. ThoughtTreasure contains 70,000 lines of C code implementing:

text agents for recognizing words, phrases, names, and phone numbers, mechanisms for learning new words, a syntactic parser, a natural language generator, a semantic parser for producing a surface-level understanding of a sentence, an anaphoric parser for resolving pronouns, planning agents for achieving goals on behalf of simulated actors, and understanding agents for producing a more detailed understanding of a discourse.

How does one extend ThoughtTreasure?When the information necessary for your application is not yet in ThoughtTreasure, you can add it!

1.4 Newell and Simon’s definition of intelligence

According to Newell and Simon ( 1976), intelligent activity is achieved through the use of- symbol patterns to represent significant aspects of a problem domain- operations on these patterns to generate potential solutions to problems-search to select a solution from among these possibilities.

1.5 Physical Symbol System hypothesis

The necessary and sufficient condition for a physical system to exhibit general intelligent action is that it be a physical symbol system

1.6 Church-Turing thesis

3 of 19

http://www.signiform.com/tt/htm/overview.htm


Computers are capable of implementing any effectively described symbolic process.

2 Machine learning

Herbert Simon defines learning as

'any change in a system that allows it to perform better the second time on repetition of the same task or on another task drawn from the same population'.

Machine learning has proven to be a fruitful area of research, producing a number of learning algorithms which vary in their goals, available training data, learning strategies and representation languages used.

3 Knowledge based systems

An expert system is a knowledge-based program that provides "expert quality" solutions to problems in a specific domain. They are not cognitive modelling programs but rather practical programs that use heuristic strategies developed by humans to solve specific classes of problems.

Many expert systems exist today. One such system to be developed is the Astronaut Science Advisor (ASA) which was used on the October 1993 flight of the Columbia space shuttle. This is a general purpose system to help astronauts work more efficiently and improve the quality of collected data. ASA tracked the time spent on each experiment and suggested ways to shorten procedures when sessions fell behind schedule and could propose a new sequence of steps to get the best and most data in the time remaining. It could also lead someone through step by step troubleshooting.

The heuristic nature of expert problem solving creates problems in the evaluation of program performance. We know that heuristic methods will sometimes fail but it is not clear exactly how often a program must be correct to be accepted. Perhaps a variation of the Turing test would be appropriate in evaluating the performance of expert systems.

3.1 Components of a KBS

The figure below illustrates the components that make up a rule based expert system. A variety of interface styles can be used such as question and answer, natural language, menu driven or graphics.

The explanation subsystem allows the program to explain its reasoning to the user including justifications for the system's conclusions or why the system needs a particular piece of data. Many systems also include a knowledge base editor.

The knowledge base contains the problem-solving knowledge of the particular application which, in a rule based system, is represented in the form of if .. then rules eg if the engine does not turn over and the lights come on then the problem is the starter motor.

The inference engine applies the knowledge to the solution of actual problems. Notice the separation of the control and knowledge elements. With this arrangement the same control and interface software can be used in a variety of systems. The expert system shell, indicated by the dashed line has no domain knowledge and such shells are commercially available for users to input their particular domain knowledge. It is important that the correct shell is chosen.

4 of 19


Architecture of an expert systemArchitecture of an expert system

User

Interface Editor

Inference

Explanation

KB

3.2 How does the inference engine work? ( for production systems)

The simplest class of expert system architectures which uses rule based representation are called production systems. A production system has three components.

A rule base - consisting of a set of production rules. A production rule has the form IF < condition> THEN < action >. Rules may have multiple conditions or actions. Rules are independent, they do not receive or pass information directly to other rules.

Working memory - containing a description of the current state of the world in a reasoning process. This description is a pattern that is matched against the condition part of a production rule to select the appropriate problem solving actions. When the condition element of a rule is matched by the contents of working memory, the action associated with that condition may then be performed.

Inference engine - which is the recognise-act cycle of the production system. The current state of the problem solving process is maintained as a set of patterns in working memory which is initialised with the beginning problem description. The patterns in working memory are matched against the conditions of the production rules which produces a subset of the rules, called the conflict set, whose conditions match the patterns in working memory. One of the rules in the conflict set is then selected ( conflict resolution ) and the rule is fired ie the action of the rule is performed, changing the contents of working memory. Then the control cycle repeats with the modified working memory. Conflict resolution strategies include refraction, recency and specificity.

3.3 Control of search in production systems

Data driven search begins with a problem description and new knowledge is inferred until a goal is reached. See below for an example.

5 of 19


# Working memory Conflict set Rule fired

0 start 6 6

1 start,v,r,q 6,5 5

2 start,v,r,q,s 6,5,2 2

3 start,v,r,q,s,p 6,5,2,1 1

4 start,v,r,q,s,p,goal 6,5,2,1 halt

1. p q goal 2. r s p 3. w r q4. t u q 5. v s 6. start v r q

Data driven search in a production system

Goal driven search begins with a goal and works backward to establish its truth. To implement this in a production system, the goal is placed in working memory and matched against the actions of the production rules. All rules whose actions or conclusions match the goal form the conflict set. When the action of a rule is matched, the conditions are added to working memory and become the new subgoals of the search. The new subgoals are then matched to the actions of other rules. The process continues until a fact is found in the initial description or by directly asking the user for specific information. The search stops when the conditions of all the rules fired in this backward fashion are found to be true. These conditions and the chain of rule firings leading to the original goal form a proof of its truth. See next page for an example of goal driven search.

Other types of expert systems exist in addition to rule based ones.

3.4 When to go for an expert system solution.

Expert systems involve a considerable investment in money and effort. Many problems are too complex or poorly understood to be suitable for an expert system.

1. Does the need for the solution justify the cost and effort of building the expert system?

2. Is human expertise available in all situations where it is needed?

3. Can the problem be solved using symbolic reasoning techniques alone?

4. Is the problem domain well structured and not require commonsense reasoning?

5. Can the problem not be solved using traditional computing techniques?

6. Do co-operative and articulate experts exist?

7. Is the problem of the proper size and scope?

6 of 19


As a rule of thumb, problems that require days or weeks for human experts to solve are probably too complex for current expert system technology. If any answer to the above questions is no then building an expert system may not be the best way forward.

# Working memory Conflict set Rule fired

0 goal 1 1

1 goal,p,q 1,2,3,4 2

2 goal,p,q,r,s 1,2,3,4,5 3

3 goal,p,q,r,s,w 1,2,3,4,5 4

4 goal,p,q,r,s,w,t,u 1,2,3,4,5 5

1. p q goal 2. r s p 3. w r q4. t u q 5. v s 6. start v r q

Goal driven search in a production system

5 goal,p,q,r,s,w,t,u,v 1,2,3,4,5,6 6

6 goal,p,q,r,s,w,t,u,v,start 1,2,3,4,5,6 halt

Expert systems differ from conventional systems in several ways.

They use knowledge rather than data to control the problem solving. Much of the knowledge used can be heuristic rather than algorithmic.

The knowledge is coded and maintained separately from the control part.

They are capable of explaining how a particular conclusion was reached and why requested information is needed during a consultation. This gives the user a chance to understand the system's reasoning, thus increasing the user's confidence in the system.

They use symbolic representations for knowledge (rules, networks, frames) and perform their inferencing through symbolic computations that closely resemble manipulations of natural language. ( An exception is neural networks.)

3.5 Other ways of representing knowledge

Mylopoulos and Levesque (1984) have classified the various representational schemes into four categories:

Logical schemes. This class uses expressions in formal logic to represent a knowledge base. First order predicate calculus is the most widely used scheme. PROLOG is an ideal language for implementing this type of scheme.

Procedural schemes. These schemes represent knowledge as a set of instructions for solving a problem. In a rule based system an IF.. THEN rule may be interpreted as a procedure for solving a goal in a problem domain. A production system may be seen as an example of this type of scheme.

7 of 19


Network schemes. These capture knowledge as a graph in which the nodes represent objects or concepts and the arcs represent relations or associations between them. Examples of this type of scheme include semantic networks, conceptual dependencies and conceptual graphs.

Structured schemes. These extend networks by allowing each node to be a complex data structure consisting of named slots with attached values. These values may be simple numeric or symbolic data, pointers to other frames or procedures for performing a particular task. Examples of this type of scheme include scripts, frames and objects.

A number of questions must be asked of any representational scheme.

Exactly what things can be represented by the objects and relations in the scheme? So if a predicate relates an object to a value eg has colour(car, red), how can that predicate be used to represent "Bill's car is redder than John's" ? How should this be captured?

Granularity of the representation. In predicate calculus and semantic networks we can denote objects in the domain only by using simple symbols. Other schemes such as frames and objects allow complex structures to be defined by encapsulating multiple features and attributes to indicate a single object in the domain.

Can the scheme distinguish between intensional and extensional knowledge? The extension of a concept is the set of all things denoted by a given concept. The intension determines what a concept means in the abstract.

How can meta-knowledge be represented?

Is inheritance provided? Another important feature of knowledge is its organisation into class hierarchies. The ability to represent the relationship between an object and its class or between a class and its superclass has proved so useful that many representation schemes include mechanisms that define these relationships.

Inheritance is a relation by which an individual assumes the properties of its class and by which properties of a class are passed on to its subclass. Eg if chimpanzee is a subclass of primate and Bonzo is a chimpanzee, we may assume that Bonzo has all the properties of chimpanzees and primates. Inheritance guarantees that all members of a class inherit the appropriate properties, ensuring consistency with the class definition. It reduces the size of the knowledge base as properties are defined once for the most general type that shares these properties rather than for all subclasses or individuals. Inheritance is also used to implement default values and exceptions.

It is often helpful to attach procedures or demons to object descriptions which are executed when an object is changed or created and provide a vehicle for implementing graphics I/O, consistency checks and interactions between objects.

3.6 Semantic networks

These are intuitive and simple to understand. They represent relations between items in a domain. The items are represented as nodes and are connected by arcs labelled with the particular relation linking the two nodes. An example net is shown below. The net serves as data which only becomes knowledge in association with a reasoning capability. Reasoning takes the form of traversing arcs in the net to identify complex relationships between the nodes. A common way of retrieving data from nets is the intersection search. A search can be made by activating all the arcs connected to one component and all the arcs connected to another. The nodes connected by these arcs can themselves be activated etc. Eventually all the paths between the components should be found and we will have found all the relations encoded in the net that link these two elements.

8 of 19


Semantic NetworksSemantic Networks

Fred

canary

isa

bird

isa

feathershas

One problem with semantic nets is the difficulty of picking the right set of semantic primitives at the beginning. A second problem is the difficulty of expressing knowledge such as

some PCs have two floppy drives. Bill does not know that George Orwell was Eric Blair.

3.7 Frames and scriptsFrames and scripts can be thought of as extensions to semantic nets. Nodes are replaced by more structured groupings of information called frames. Frames are able to represent class/subclass dependencies. The expressive power of simple frame systems lies in the fact that

each frame is a fairly detailed description of what an object is. Frames can also embody procedural knowledge which is done by procedural attachment to slots.

information can be transferred between them by inheritance.

Each individual frame may be seen as a data structure, similar in many respects to the 'record', that contains information such as

frame identification relationship to other frames descriptors of requirements for frame match procedural information on the use of the structure described frame default information new instance information.

The presence, absence or amount of detail in these slots depend on the particular problem being addressed.

Frames extend semantic nets in a number of ways. It makes it easier to know what is being described and to organise the knowledge hierarchically. It may be useful to think of an object as a single entity for some purposes and only consider details of its internal structure for other purposes.

Procedural attachment is an important feature of frames as well as supporting class inheritance. The slots and default values of a class frame are inherited across the class/subclass and class/member hierarchy. When an instance of the

9 of 19


class frame is created, the system will attempt to fill its slots, either by querying the user, accepting the default value from the class frame or executing some procedure or demon to obtain the instance value.

ScriptsA script is a structured representation describing a stereotyped sequence of events in a particular context, particularly in natural language understanding.

Other issues involved with knowledge representation

Selection and granularity of symbols for the knowledge base. Exhaustiveness. If a representation is exhaustive, there will be no unspecified side effects and the frame problem

effectively disappears. Plasticity Computational efficiency

3.8 Assessing an expert system

Quality can be considered in terms of technical correctness and business correctness. Criteria contributing to technical correctness include

usability security efficiency reliability maintainability adaptability portability.

Criteria contributing to business correctness include

timeliness cost/benefit ease of transition userfriendliness.

3.9 Tools available to build expert systems

These may be classed into the following categories:

expert system shells development environments such as Kappa symbolic languages such as Prolog and Lisp algorithmic languages

4 The knowledge engineering processThe primary people involved in building an expert system are the knowledge engineer, domain expert and the end user.

The knowledge engineer's job is to extract the necessary knowledge from the domain expert, choose the correct software and hardware tools as well as implementing that knowledge in a correct and efficient knowledge base. He/she may initially be ignorant of the application domain. The skills of a knowledge engineer should include fast learner, effective communicator, knowledge of a variety of acquisition techniques, knowledge base design experience, organised good recordkeeper , conceptualises well, knowledge in many diverse areas

The domain expert provides the knowledge of the problem area and is usually someone who has worked in the domain area and understands its problem solving techniques eg using shortcuts, handling imprecise data, evaluating partial solutions and all the other skills that mark a person as an expert. He/she is primarily responsible for spelling out these skills to the knowledge engineer.

10 of 19


4.1 Concerns of the expert

How much time and effort is required?Is co-operation with the project compulsory?Who will take on my current workload?Will my weaknesses be exposed?How will this project affect my relationship with colleagues and management?Will confidentiality be respected?What use will be made of audio/visual recordings?

The skills and needs of the user must be considered throughout the development process.

What level of explanation does the user need?Can the user provide correct information to the system?Is the user interface appropriate?Will the system make the user's work easier, quicker or more comfortable?

Generally work on the system begins with the knowledge engineer trying to gain familiarity with the problem domain. This helps in communicating with the domain expert. Next the knowledge engineer and the expert begin the process of trying to extract the expert's problem solving knowledge. Once the knowledge engineer has obtained a general overview of the problem domain and gone through several problem solving sessions with the domain expert, he/she is ready to begin actual design of the system - selecting a way to represent the knowledge, determining the search strategy and the user interface. Usually a prototype is built to provide a test bed for the preliminary design assumptions. The knowledge gained can be refined, added to or modified as appropriate.

4.2 Knowledge Acquisition

Buchanan et al present a framework for knowledge acquisition that identifies the major stages as (see diagram)

Knowledge acquisitionKnowledge acquisition

Identification

Conceptualisation

Formalisation

Implementation

Testing

Identification - problem characteristics are identified.

11 of 19


Conceptualisation - how primary concepts and key relationships among the concepts in the domain are depicted and related by the domain experts.

Formalisation - requires the knowledge engineer to map the recognised concepts and relations into formal representation mechanisms.

Implementation

Testing

4.3 What is knowledge?

A knowledge engineer has to conceptualise what a domain expert knows and uses to solve problems. So what is knowledge?

Feigenbaum (1983) suggests that Knowledge is information that has been interpreted, categorised, applied and revised.

Waterman (1981) suggests that Knowledge can be exemplified by concepts, constraints, heuristic methods for probabilistic data and principles that govern domain specific operations.

Hayes-Roth (1983) suggests thatDomain knowledge consists of descriptions, relationships and procedures.

Humans use concepts to create their symbolic understanding of the world. A concept is a symbol that stands for a common characteristic or relationship shared by objects or events that are otherwise different (Kagan 1972). Salient features are those that distinguish the objects, events or ideas that the concept represents from other concepts. Consider sugar and salt. To describe them we focus on colour, size, function and taste. Background and context would influence which features would be the most salient eg a chemist might consider their chemical properties whereas a chef might consider their function as being the most salient feature.

A concept hierarchy is a structural taxonomy or arrangement of the associations that make up a concept. The strongest associations are recalled immediately. The concepts we have learned, the way we have structured them into hierarchies and our system for interweaving them comprise our personal method of organising our worlds. This organisation influences what we choose to attend to, how we perceive inputs and how we organise what we sense. Consequently, it influences the thinking patterns, judgement and perception of the domain experts with whom knowledge engineers work.

4.4 Knowledge and expertise

We cannot directly observe someone's knowledge but we can observe and identify expertise. If two individuals possess the same knowledge will they apply their knowledge in the same way? The answer is no and the difference is the notion of expertise.

Experts are more knowledgeable about their domain and can apply and use knowledge more effectively than novices. Experience, not innate capacity, seems to be the major differentiator between novices and experts. It influences knowledge organisation, storage and retrieval. Experience is the factor that changes unrelated facts into expertknowledge. (Kolodner 1983)

Chi, Glaser and Rees (1981) suggest that the schema a novice uses represents surface features of a problem while the schema an expert uses represents the more abstract semantic principles. Superior organisation of knowledge seems to be critical to expert performance. Experts not only have more knowledge but they organise it into more meaningful "chunks". Chunks are groups of items that are stored and recalled together. Although chunking is useful to experts, it may inhibit their ability to be cognisant of their own knowledge and express it to others. Performance over time becomes more automatic and hence requires less cognitive awareness.

12 of 19


4.5 Types of knowledge

Procedural knowledge ( knowing how)This includes the skills an individual knows how to perform eg executing a sharp turn while riding a bicycle. Knowledge of one's mother tongue is procedural - we possess this knowledge but find it very difficult to state rules that describe precisely how to use the language. Addis (1993) suggests that this type of knowledge may be divided intothree types

kinaesthetic knowledge perceptual knowledge gestalt

The procedures for carrying out these kinds of skills are deeply embedded and linked sequentially.

Declarative knowledge ( knowing that)This represents surface level information that experts can verbalise. Consider how difficult it would be to verbally describe how to ride a bicycle ( procedural knowledge). Declarative knowledge is useful in the initial stages of knowledge acquisition but is of less value in later stages. It is an expression of what the expert is aware of knowing.

Semantic knowledgeThis reflects cognitive structure, organisation and representation. It has been described as organised knowledge about

words and other verbal symbols word/symbol meanings and usage rules word/symbol referents and interrelationships algorithms for manipulating symbols, concepts, relations.

Episodic knowledgeThis is autobiographical, experiential information that the expert has grouped or chunked by episodes. It contains information about " temporally dated episodes or events and temporal-spatial relations among these events". It is believed to be organised by time and place of occurrence. An example of episodic knowledge would be 'driving to work'. 4.6 Techniques for knowledge elicitation

Much has been written about techniques for knowledge elicitation. We will simply mention some of the techniques without elaborating in great detail.

InterviewingThis technique has several variants

unstructured interview structured interview focused interview forward scenario simulation

Protocol analysisThere are different types of protocol analysis

behavioural verbal

Multi-dimensional techniquesThese include

card sorting repertory grids

Goal decomposition techniquesThese include

13 of 19


laddered grid twenty questions

5. RoboCup poses a new set of AI research challenges

Artificial Intelligence is an umbrella term which covers areas such as robotics, vision, natural language processing, machine learning, knowledge based systems.

With the defeat of world chess champion Garry Kasparov by IBM's Deep Blue machine, one of AI’s long standing research challenges was at last achieved. 40 years of innovation by many great researchers prepared the way for Deep Blue's crowning achievement. Now a new AI challenge is on the horizon: RoboCup, the robot world soccer cup. ( www.robocup.org )

5.1 The RoboCup challenge

RoboCup is intended to stimulate innovations in a wide range of technologies and the integration of technologies into a fully functional robotic soccer team. Given how primitive today's robots are compared to a good soccer team of 11 players, the inherent technical challenges are plentiful. To build a world-class roboteam will require

capable single players (autonomous agents, robotics, vision, and real-time sensor fusion), teamwork (multiagent collaboration, context recognition), understanding the competition (cognitive modeling), the ability to develop and execute plays and strategies in real time (strategy acquisition, real-time reasoning

and planning, and reactive behaviour), and pre- and post-game training (machine learning),

to name a few items of necessary functionality.

RoboCup-97, held in conjunction with the 1997 International Joint Conference on AI, in Japan, was a preliminary step toward meeting this challenge. For the most part, play level was more akin to that of first and second graders learning how to play than that of a proficient (let alone world-class) college team. "Chess took us 40 years," notes Hiroaki Kitano, founder of RoboCup and a senior researcher at the Sony Computer Science Lab. "With real robots, it will take us at least that long or longer."

5.2 Three different leagues

RoboCup 97 featured three different leagues. The Middle-Size and Small Leagues involved physical robots; the Sim-ulation League was for virtual, synthetic teams. RoboCup organizers posited three specific challenges for the physical robots:

moving the ball to the specified area (shooting, passing, and dribbling) with no stationary, or moving obstacles; catching the ball from an opponent or a teammate (receiving, goal-keeping and intercepting), and passing the ball between two players.

The first two involve individual agent skills, and the third introduces simple cooperative behaviour. All are appropriate for beginner-level soccer.

Likewise, for the Simulation League, RoboCup issued a three-part Synthetic Agents Challenge '97, including

Learning: offline skill learning by individual agents, offline collaborative learning by teams of agents, online skill and collaborative learning, and online adversarial learning. Teamwork: contingency planning for multiagent adversarial game playing, plan decomposition and merging, and executing team plans. Opponent modeling: online tracking of opponents' behaviour and intentions, online strategy recognition by sideline coach agents, and offline review after the game.

5.3 Limited results in 1997

These are lofty challenges indeed. In reality, the physical robots that competed were still quite brittle. They were sensitive to even the slightest changes in lighting and colour conditions from their practice fields at home to those in

14 of 19

http://www.robocup.org/


the Japanese competition field. Many bots had difficulty even finding the ball, and spent a good deal of time just standing around looking for it. Robots ran into each other, pushing each other around. Indeed, all the pushing led to some robots starting to smoke as motors overheated and began to melt plastic chassis.

The pace of play was more akin to walking, or even handicapped walking, than running. Nor did the bots have much agility. They had no appendage with which to kick the ball. All bots simply pushed the ball along. Plus, the teams in the physical leagues showed very limited teamwork. There was much more sophisticated play in the Simulation League - an environment where researchers could focus on some of the higher-level tasks without worrying about lower-level issues such as robot controls and short battery .

5.4 Rules of the game

RoboCup-97 had modified soccer rules. Players received no penalties for being offside, for example. The physical robot leagues allowed a maximum of only five robots per team. Middle-size robots had to be within 50 centimetres in diameter, but there was no height restriction. Weight ranged from 100 pounds to the winning DreamTeam robots, which were less than a foot high and weighed under two pounds each. Small robots could be as large as 15 centimetres in diameter.The middle-size field was roughly the size of a tennis court, and the small field was a Ping-Pong table. Both had walls. Middle-size robots used a size 4 soccer ball painted red, and the small bots an orange golf ball. For the Middle-Size League, the game consisted of a five-minute first half. a 10-minute break, and a five-minute second half. For the Small League, each half and the break were 10 minutes.RoboCup provided the soccer simulator for the Simulation League. A full 11 players were on a simulation team. The simulator is available at the RoboCup Web site. It was designed with lots of obstacles to make it more realistic. For instance, it was difficult to figure out where a player was, and if a player ran fast, it got tired.

5.5 Middle-Size League

Although the Middle-Size League comprised fewer than half a dozen competitors, the teams exhibited a diversity of technologies. The cowinners, for instance, were on opposite ends of the robot-control spectrum. The DreamTeam from the University of Southern California's Information Sciences Institute had autonomous robots, each with an on-board camera and computer. The robots from Osaka University had cameras on board, but they were controlled by a central computer that issued commands to the robots. Despite problems with radio interference, the Osaka team managed to tie for first.The team from USC/ISI built and programmed their team of small, lightweight, autonomous agents in 6 months. Each robot consisted of a small 80x86-based motherboard, camera, and batteries on board a four-wheel-drive model car. The camera was mounted up front, with a plastic bumper below for pushing the ball."Our autonomous agent architecture was key to our win," according to DreamTeam developer Rogelio Adobbati, a research assistant at USC/ISI. This included a vision module to process visual input, a drive controller to steer the bot and a decision engine to help the bot decide what to do.The decision engine had several components: a model manager to keep track of the field and nearby objects, a strategy planner to react to situations appropriately, and specialized components for goalkeeper, forward, and defender positions. The strategy was hard-coded, using C++.

5.6 Small League

CMUnited from Carnegie Mellon University won this league. The developers used machine-learning techniques and achieved some multiagent innovations in team formations, position switching, and passing. The team distinguished itself from many of the physical teams in both leagues by being able to pass and kick the ball and by showing some team-work. For example, at least two of the goals the team scored involved one robot passing the ball to another, which shot the ball into the goal."We chose the Small League to test a full version of our team in limited lab space using an overhead camera, and to focus on strategy and AI innovations instead of low-level control issues," explains Stone.The CMUnited team included five robots, plus one on the bench that was used when a first-string player's batteries were low. Each bot was 12 cubic centimetres arid weighed about two pounds, with approximately 60% of the weight being the batteries. With a single overhead camera, image processing was done off board and relayed via a wire less radio link. The vision system was able to determine both the position and orientation of 10 robots, and the ball position and velocity.As with any team, much of the success was in the training and preparation. "We carefully tested situations that would come up repeatedly, such as a ball moving towards the goalkeeper. We threw the ball towards the goalie over and over again until we were satisfied with the performance," says Stone. "We also worked from the bottom up, first making sure that the robots could hit a stationary ball into an empty goal, then working on moving balls, and finally reasoning about the positions of teammates and opponents. The reasoning was accomplished with evaluation func-tions that determined a free path from the ball to the goal."

15 of 19


Why did CMUnited do so well? "Our vision system, unlike any other, was almost noiseless," according to Stone. "That was a big advantage. We also had very good strategic teamwork: the players were able to switch positions and pass the ball to each other. We were the only team to meet the challenge of getting five robots to work as a team in a noisy, fast environment. The biggest need for improvement is in robot speed. They actually moved fairly slowly (a walking pace). We could have moved them faster," he continues, "but control and reliability would have suffered. We were focused on exhibiting reliable, intelligent behaviours."

5.7 Freed from physical bodies

CMUnited also competed in the Simulation League (they came in fourth losing in the semifinals to the eventual champions) Stone found distinct advantages in the simulator. "With real robots, we spent most of our time just getting the sensors and actuators to work," he explains. "Only at the end were we able to focus on strategy. However, the simulator abstracts away the perception and action programs. Therefore it is possible to focus on more high-level issues. For that reason, my research on layered learning is going on mostly in the simulator."

Stone's layered-learning research is his approach to building multiagent systems using machine learning. Much as a child learns to play soccer, this approach layers increasingly complex learned behaviours. In the world of soccer, for instance, first a player (agent) learns low-level skills to control the ball. Then, building on this learned skill, it learns the higher-level skills of playing with teammates (multiple cooperating agents).

To train the virtual player, Stone's group provided the agent with a large number of training examples and used neural networks. Once the player could control the ball, it was ready to be out on the field playing with the team. Passing the ball from one player to another is a fundamental component of team play. The developers successfully employed decision trees to help the player decide whether or not to pass.

5.8 Simulation League

The CMUnited team effort illustrates the intent of the RoboCup organizers to push and test new technologies. For USC/ISI researchers already deeply involved in multiagent collaboration work for the mili tary, RoboCup offered a different testbed for their innovations. The Soar system developed at USC/ISI has been used for some time to build synthetic agents for military exercises. Milind Tambe, a research scientist at USC/ISI, has been working on building a layer on top of Soar for multiagent collaboration arid teamwork. Approximately 30% of the code was general enough to transfer directly from a military application to the RoboCup soccer domain.

Their soccer team, ISIS (ISI Synthetic), was the top US team in the Simulation League, coming in third in the competition. Tambe is already looking toward next year. "This year we defended well. Our approach worked well," asserts Tambe. 'As for next year, we need a structured offensive strategy. We need better agent modeling and plan-recognition abilities to understand what our opponents are up to."Humboldt was the winning team for the Simulation League. Harking from Humboldt University in Germany, the developers used agent-oriented programming in the belief-desire-intention style to design the virtual team. As team leader Hans-Dieter Burkhard describes their approach, agents (players) have beliefs about the world that are updated according to incoming sensor information. The agents then compute the coordinates of objects. The agents also have action sequences for intercepting the ball, reaching certain positions on the field, and manipulating arid kicking the ball.The Humboldt agents also have a desire/goal component. Based on their beliefs, the players decide which desire to adopt (such as intercept, kick, dribble, or run). Each agent is so well-articulated that it can evaluate various plans based on the world model and choose a course of action. Although the programmers hoped to use case-based reasoning to adapt playing to the opponents' behaviour, it was not ready to use this year.Given this approach, the Humboldt team has not used any machine-learning techniques. "We have learned," says Burkhard, "that learning should not start with simple behaviour like single kicks, but it should be used to improve the behaviour. First, make a careful analysis and implement a raw skill; then tune it by learning methods. For man-ipulating the ball, for instance, you need several steps that are very difficult to learn as a sequence from scratch." Once the basic skills were in place, the developers found that Al learning and planning methods worked well to tune behaviour.

5.9 Getting better

Although still quite seminal, robot soccer promises to be a challenging testbed for AI. "Already, we're seeing the beginnings of teamwork. In a domain such as soccer, where it is very difficult to program teamwork, we're seeing Al systems doing a better job than hardcoded, procedural code systems. This is encouraging to the Al community," notes Kitano."Now people are back in their offices understanding what to do for next year," he continues. "I expect very rapid progress in the next five to 10 years in this area." The challenges are set. Future visions are of legged robots,

16 of 19


humanoid robots, and a telepresence competition. Robots will cooperate in space, on terra firma, and in cyberspace in a variety of endeavours.Will there ever be a world-class match between humans and robots? Who would want to collide with a metal robot on the soccer field? Or kick it in the shins and break a foot? Many bot innovations must occur before such a day: softer materials for robots, a different design for agility, better batteries for endurance, and so on. Remembering what Kitano said about the AI community taking 40 years to rise to the grand challenge in chess, we'd be wise to adopt an evolutionary view of the field.

5.10 The Future

In 1998 a new league was introduced – Sony Legged Robot League. The 4 legged robot has a total of 16 degrees of freedom (dof): each leg has 3 joints, the head has pan, tilt and roll movements and the tail has 1 dof .

Before actually playing soccer with human players, RoboCup will organise humanoid leagues in the following categories and start the Humanoid league from 2002.Fully autonomous humanoid league: soccer games by teams of fully autonomous humanoid biped robots. A regular league will be performed by humanoid robots of height equivalent to a real human.Tele-operational humanoid league: the operator is allowed to control the robot only through the information obtained by sensors on board the robot.Virtual humanoid league: soccer with teams of simulated humanoid robots, with high quality computer graphics, accurate physics simulation, vision and sensor simulation.

Knowledge Elicitation Exercise

Extract From Transcript Of Interview With Geologist

E = Expert, K = Knowledge Elicitor

E These are all fine grained, these are all volcanic rocks, yeah. Now the difference here is going to be in, you can’t really identify the minerals very easily cause they’re fine grained and the minerals could be only a twentieth of a millimetre across in these. What you’re going to be looking for is colour, density, general sort

17 of 19


of, the way it looks, whether it’s very light or whether it’s sort of getting half way. Andesite is a volcanic equivalent of diorite in which case it could have a bit of quartz in, otherwise it’s a mixture of feldspars and a few pyroxenes, perhaps the odd amphibole in there. It will be generally be quite dark, because despite the fact that it’s got quite a lot of light minerals in, when you mix minerals up, when you get a lot of very fine grained material darker material tends to show more than lighter material so andesite will still look quite dark. But it won’t be as dense as some of the basalts for instance. And you might be able to pick out with a hand lens the grains of feldspar. And of course if it’s porphyritic, that is if it contains these unusual large crystals it might have some feldspars in it which will give it away.

Then you’ve got two pairs left, really. Perhaps it’s easiest to deal with the rhyolite first. The rhyolite is the volcanic equivalent of a granite. Very very light, low density. very very light in colour. And the other thing that rhyolite is, because it’s got more silica in, because of structural reasons, it forms a more viscous lava. Basalt lava is very runny, rhyolite lava is very sticky and so you’re likely to get flow textures preserved in rhyolite, which you don’t in basalt. You’re likely to get lineations of minerals, almost as though you could see the streams of the crystals going through. So the rhyolite might give you some idea, you might see this flowage, this sort of treacly flow.

Er.. now trachyte, this is where it gets difficult because they’re fine grained and they’ll have the same density, I mean, rhyodacite, dacite and trachyte, they’re all very difficult to tell apart in hand specimen. Trachyte is more alkaline, it’ll contain more of these alkaline minerals, particularly potassium rich feldspars, possibly the odd feldspathoid, which is a feldspar like mineral only with less silica than a normal feldspar. Might have a few of those in. The one way you might be able to tell trachyte is that it has this peculiar texture in which is trachytic flow texture in which you get the feldspars lined up like a series of dominoes across the rock. And what’s happened is that the feldspars have crystallised and then been dragged out by the flow of the fluid. It’s called trachytic flow texture and its often quite useful. It’s distinctive from rhyolitic flow texture, which is more of a banded appearance rather than this distinct flow of individual grains. But in general trachyte is more alkaline than these two. That’s going to be difficult to tell in hand specimen cause you can’t identify the minerals so well.

Dacite and rhyodacite, er, dacite, the difference between dacite and rhyolite is it’s got less silica, so it’ll have less quartz in, might have the odd pyroxene, perhaps a few more amphiboles. Quite a lot of plagioclase feldspar with the odd alkali feldspar. Rhydocite is obviously half way between the two, very difficult to tell the difference between the two, very difficult indeed. You’d need chemical analysis, really, I think.

K Er..you’ve mentioned two textures. What other textures are there that you haven’t mentioned so far?

E Quite a few. I’ve mentioned one, the porphyritic texture which I mentioned. That is the case where you’ve got large crystals seemingly alien crystals, in a fine grained ground mass or matrix. What these represent is, they’re crystals which have solidified at depth, then the lava has come onto the surface and most of it has crystalised into this very fine grained stuff, but you’ve still got a few of these phenocrysts left there. That’s one, that’s porphyritic texture.

K Which of those rocks do you tend to find that one in?

E Well, the porphoritic texture could apply to any of the volcanics but in general you tend to find them in, I would say that you’re more likely to get basalts and andesites as being porphortic rather than rhyodites and dacites simply because the minerals which produce phenocrytes are more likely to cool at the same, a different temperature to the general magma of basalt and andesite. Porphoritic rhyolites and dacites, trachytes aren’t so common. Basalts cool at about three hundred degrees higher sorry, crystalise at about three hundred degrees higher than rhyolites. So if the phenocryst cools somewhere in between there you might not get them in one or the other, it depends on the cooling history. In general, the more silica poor rocks have more phenocrytes.

K And er, what other textures are there then after those three?

E Well, textures really relate to the way in which minerals interlock. You can get some cases where the minerals enclose other minerals, you might have a large grain of one mineral. That’s called ophitic texture - o p h t i c - (spelled out by subject). Or in the case where minerals are partially enclosed it’s called poiklitic texture.

K Could you give us an example of that, do you think?

E Well, ophitic and poiklitic textures you might find in gabbors, diorites; in general these textures are better, you can pick them out better in the coarser grained rocks, although the phenocrysts may show them in the finer grained rocks. Er, graphitic texture whereby minerals have intergrown at the same temperature to

18 of 19


produce a very, a very very, er, easy to identify texture, it’s sort of straight intergrowths, looks very, er, sort of blocky. They commonly occur in deep seated rocks. You can get peculiar intergrowths such as myrmekites, which are sort of like worm like intercollations of minerals where the minerals haven’t they’re so mixed up and they’ve intergrown together at the same temperature they haven’t been able to form their normal Crystal shape, they just form these swirling intergrowths.

K Could you give us some more examples of graphitic texture from among the sample?

E Graphitic texture. I think you’re most - your tape’s finished.

K Right, could you give us some examples of graphitic texture?

E Right. Er, you may get graphitic texture in the intergrowth of plagioclase and quartz, for instance. They cool at relatively low temperatures, er, I think quartz crystallises at about seven hundred degrees centigrade, plagioclase slightly higher, but in , if the conditions are right, you may get intergrowths of plagioclase and pyroxene in a similar fashion in a gabbro.

K Could you give us some examples of poiklitic textures?

E Yeah, poiklitic textures you might, you could find poiklitic textures in any of these, really, particularly I think minerals containing both plagioclase and pyroxenes, you’re more likely to get them in the gabros, diorites, basalts. You get one, you get lots of little crystals of perhaps pyroxene laths, and then later in the cooling history a big plagioclase crystal grows round them, partially enclosing them, poiklitic texture. That’s quite a common occurrence between feldspar and pyroxene, so you’re likely to get that in the ones containing those. Such as basalts, gabbros, things like that.

19 of 19

Documents

Intelligence and Knowledge