Artificial Intelligence

Basics

Univ.-Lektor Dr.techn. Alexander K. Seewald

Alexander K. Seewald 1
alex@seewald.at / alex.seewald.at
 Spam Filtering
Problem
• Spam : Nonspam = 17 : 1; 200 spams/day

Solution: State-of-the-Art System

• Bayesian filter - provably better than human
• Deletes 99.8% of spam
• Few nonspam mails deleted
(<0.1%)
• Low maintenance, working
towards zero maintenance

Alexander K. Seewald 2
alex@seewald.at / alex.seewald.at
 Stahlwerk Bous & Siemens

• Optimization of melting process with NN and

analytical model: Steel production +6,0%; Energy
consumption -3,1%
Alexander K. Seewald 3
alex@seewald.at / alex.seewald.at
RoboSail Systems
• Autopilot for one-person
sailing
• Race-proven with
various state-of-the-art
AI and ML components.
• Human jargon like gust,
close-hauled, luff as
background knowledge!

Alexander K. Seewald 4
alex@seewald.at / alex.seewald.at
 What is Artificial Intelligence?
Systems that think like humans Systems that think rationally
"The exciting new effort to make "The study of mental faculties
computers think... machines with through the use of computational
minds, in the full and literal sense" models" (Charniak and McDermott,
(Haugeland, 1985) 1985)
"[The automation of] activities that
we associate with human thinking, "The study of the computations that
activities such as decision-making, make it possible to perceive, reason
problem solving, learning..." and act" (Winston, 1992)
(Bellman, 1978)
Systems that act like humans Systems that act rationally
"The art of creating machines that "A field of study that seeks to
perform functions that require explain and emulate intelligent
intelligence when performed by behavior in terms of computational
people" (Kurzweil, 1990) processes" (Schalkoff, 1990)

"The study of how to make "The branch of computer science

computers do thinks at which, at the that is concerned with the
moment, people are better" (Rich automation of intelligent behavior"
and Knight, 1991) (Luger and Stubblefield, 1993)

Alexander K. Seewald
5
alex@seewald.at /
 Systems that act like humans
The Turing Test
Computing machinery and intelligence [Turing, 1950]
• "Can machines think?" ⇒ "Can machines behave
intelligently?"
• Operational test for intelligent behavior = Imitation Game

• Pred. 30% chance for machine to fool lay person for 5mins
• Anticipated all major arguments against AI(!)
• Suggested major components of AI: knowledge, reasoning,
language understanding, learning

Problem: Turing test is not reproducible and not constructive.

Alexander K. Seewald 6
alex@seewald.at / alex.seewald.at
 Systems that think like humans
Cognitive Science
1960s Cognitive Revolution: information processing
psychology replaced prevailing orthodoxy of behaviourism

Requires scientific theories of brain's internal activities

• Abstraction - level of Knowledge, Assemblies, Neurons...
• Validation - requires predicting and testing behavior of
human subjects (top-down = Cognitive Science); and direct
identification from neurological data (bottom-up =
Cognitive Neuroscience)

Both approaches are distinct from AI; but still share direction.

Alexander K. Seewald 7
alex@seewald.at / alex.seewald.at
 Systems that think rationally
Laws of Thought
• Normative (or prescriptive) rather than descriptive.
• Aristotle: what are correct arguments / thought processes?
• Several Greek schools developed various forms of logic =
notation and rules of derivation for thoughts; may or may
not have proceeded to the idea of mechanization.

• Direct line via mathematics and philosophy to modern AI

Problems
• Not all intelligent behavior is related to logical deliberation
• The purpose of thinking = What thoughts should I have?

Alexander K. Seewald 8
alex@seewald.at / alex.seewald.at
 Systems that act rationally
Doing the right thing
• Rational behaviour: doing the right thing
• The right thing: which is expected to maximize goal
achievement given the available information
• Doesn't necessarily involve thinking, but thinking should
be in the service of rational action.

Aristotle (Nicomachean Ethics):

Every art and every inquiry, and similarly every action and
pursuit, is thought to aim at some good.

Alexander K. Seewald 9
alex@seewald.at / alex.seewald.at
 AI prehistory
Philosophy logic, methods of reasoning
mind as physical system
foundations of learning, language, rationality
Mathematics formal representation and proof
algorithms, computation, (un)decidability,
(in)tractability, probability
Psychology adaptation, phenomena of perception and
motor control, experimental techniques
Economics formal theory of rational decisions
Linguistics knowledge representation, grammar
Neuroscience plastic physical substract for mental activity
Control theory homeostatic systems, stability
simple optimal agent designs

Alexander K. Seewald 10
alex@seewald.at / alex.seewald.at
 History of AI
1943 McCulloch & Pitts: Boolean circuit model of brain
1950 Turing's Computing Machinery and Intelligence
1952-69 Look, Ma, no hands! - Phase
1950s Early AI programs: Samuel's checkers, Newell &
Simon's Logic Theorist; Winograd's Blocks World
1956 Dartmouth meeting: Artificial Intelligence adopted
1965 Robinsons complete logical reasoning algorithm
1966-74 AI discovers computational complexity
1969-79 Early development of knowledge-based systems
1980-88 Expert systems industry booms
1988-93 Expert systems industry busts: "AI Winter"
1988- Resurgence of probability; increase in technical depth
"Nouvelle AI": ALife, GAs, soft computing
1995- Agents metaphor

Alexander K. Seewald 11
alex@seewald.at / alex.seewald.at
 Agents and environments
An agent is everything that
perceives and acts.
The whole field of AI can
be viewed as being
concerned with design of
intelligent agents.
Agents include humans, robots, softbots, vacuums cleaners...
The agent function maps from percept histories to actions:
f: P* → A
For any given class of environments and tasks, we seek the
agent with the best performance. Computational limitations
make perfect rationality unachievable.

Alexander K. Seewald 12
alex@seewald.at / alex.seewald.at
 Types of agents
Four basic agent types in order of increasing generality:
• Simple reflex agent
• Reflex agent with state
• Goal-based agent
• Utility-based agent

All these can be turned into learning agents, where some

aspects of the agent can be changed by experience.
Learning is the central issue for intelligent agents. The
research fields of Machine Learning and Data Mining have
investigated simpler learning model for decades. While a
general learning agent is still decades away, ML & DM are
well on the way towards a mature field.

Alexander K. Seewald 13
alex@seewald.at / alex.seewald.at
Simple reflex agent

Alexander K. Seewald 14
alex@seewald.at / alex.seewald.at
 Example: Vacuum cleaner agent

Percepts: clean/dirty, wall, stairs

Actions: move, rotate, clean
Goals: maximize amount of dirt collected / cleanliness
Environment: single-level household

Alexander K. Seewald 15
alex@seewald.at / alex.seewald.at
Reflex agent with state

Alexander K. Seewald 16
alex@seewald.at / alex.seewald.at
 Example: Ant-based routing

[Di Caro & Dorigo, 1998] have shown that ant-based routing
outperforms other common routing methods. State is the
history of visited nodes; similar to pheromone tracks in
real ants.

Alexander K. Seewald 17
alex@seewald.at / alex.seewald.at
Goal-based agent

Alexander K. Seewald 18
alex@seewald.at / alex.seewald.at
 Example: RoboCat

RoboCat (Seewald, 1999; Diploma thesis) is an example for a

goal-based robot. The goal in that case was to follow and
hit blue objects - balls, mostly.

Alexander K. Seewald 19
alex@seewald.at / alex.seewald.at
Utility-based agent

Alexander K. Seewald 20
alex@seewald.at / alex.seewald.at
 Example: Invisible Person

sources\IP_TTT.MPG

The Invisible Person project with the Technical Museum in

Vienna was concerned with the creation of an engaging
playful agent. The agent group at ÖFAI was responsible
for modelling its behaviour.

Alexander K. Seewald 21
alex@seewald.at / alex.seewald.at
Simple Learning Agent (reflex-based)

Alexander K. Seewald 22
alex@seewald.at / alex.seewald.at
 Example: Stanley
Autonomous robot vehicle which won the DARPA Challenge
2005. Built at Stanford University in about 15 months by a
team of around 35 people. Uses Machine-Learned Laser
Perception and Speed Strategy.

Presenting Stanley

Alexander K. Seewald 23
alex@seewald.at / alex.seewald.at
 How can we build such agents?

• Search / Problem Solving

• Knowledge and Reasoning; Planning
• Acting under Uncertainty
• Decision Theory
• Communication / NLP

• Learning

Alexander K. Seewald 24
alex@seewald.at / alex.seewald.at
 Search / Problem Solving
Search is a central theme in AI. The fastest path through a
city; VLSI layout; the correct interpretation of a given
sentence; and even general learning - all these can be
formulated as search problems.

Alexander K. Seewald 25
alex@seewald.at / alex.seewald.at
 Search / Problem Solving
A problem consists of: the initial state, a set of operators, a
goal test function, and a path cost function. The
environment of the problem is represented by a state
space.

Alexander K. Seewald 26
alex@seewald.at / alex.seewald.at
 Search / Problem Solving
A single general search algorithm can be used to solve any
problem. Search algorithms are judged on completeness,
optimality, time complexity and space complexity.
Complexity depends on b, the branching factor; and d, the
depth of the shallowest solution.

Breadth-first search expands the shallowest nodes in the

search tree first. It is complete, optimal for unit-cost
operators, and has time and space complexity of O(bd).

Uniform-cost search expands the least-cost leaf node first. It

is complete, and optimal for any cost function. Its space
and time complexity is the same as Breadth-first search.

Alexander K. Seewald 27
alex@seewald.at / alex.seewald.at
 Search / Problem Solving
Depth-first search expands the deepest node in the search
tree first. It is neither complete nor optimal, and has time
complexity of O(bm) and space complexity of O(bm),
where m is the maximum depth.

Depth-limited search places a limit on the depth of depth-

first search. It is complete if the limit is greater than the
depth of the shallowest solution.

Iterative deepening search calls depth-limited search with

increasing limits until a goal is found. It is complete;
optimal for unit-cost operators, and has time complexity of
O(bd) and space complexity of O(bd). Preferred method in
large search spaces when depth of solution is not known.

Alexander K. Seewald 28
alex@seewald.at / alex.seewald.at
 Search / Problem Solving

Searching the full state-space is only feasible for very small

problems. Informed search algorithms take advantages of
heuristics to prune large portions of the search space to
improve time complexity in the average case. Worst case
time complexity is unchanged.

Best-first search expands the minimum cost node first. The

following search strategies are variants of best-first search.

Greedy search minimizes the estimated cost to reach the

goal. Search time is usually reduced, but optimality and
completeness are lost.

Alexander K. Seewald 29
alex@seewald.at / alex.seewald.at
 Search / Problem Solving
A* search minimizes the current cost plus the estimated cost
to the goal. If the latter is never overestimated (admissible
heuristic) and we handle repeated states, A* is complete,
optimal, and optimally efficient among all optimal search
algorithms for a given admissible heuristic. Its space
complexity is still exponential in problem size.

Refinements such as iterative deepening A* and simplified

memory-bounded A* address this problem.

Interestingly, some search problems are quite hard for

humans, so even our refined in-built heuristics are not
perfect.

Alexander K. Seewald 30
alex@seewald.at / alex.seewald.at
Example: A* search

Alexander K. Seewald 31
alex@seewald.at / alex.seewald.at
 Search / Problem Solving

Iterative improvement keeps only a single state in memory,

but can get stuck on local maxima. Simulated annealing
provides a way to escape local maxima, and is complete
and optimal given a long enough cooling schedule.

For constraint satisfaction problems, variable and value

ordering heuristics provide solutions very quickly even for
very large problems. Appropriate understanding and
modeling of the problem domain is essential.

Alexander K. Seewald 32
alex@seewald.at / alex.seewald.at
Example: Game as Search

Alexander K. Seewald 33
alex@seewald.at / alex.seewald.at
 Knowledge and Reasoning

Intelligent agents need knowledge about the world in order to

reach good decisions. Humans use huge amounts of
common-sense knowledge to solve even tiny tasks.

Knowledge is stored in the form of sentences in a knowledge

representation language that are stored in a knowledge
base.

A knowledge-based agent operaters by storing sentences

about the world in its knowledge base; using an inference
mechanism to infer new sentences, and using them to
decide what action to take.

Alexander K. Seewald 34
alex@seewald.at / alex.seewald.at
 Knowledge and Reasoning

A representation language is defined by its syntax and

semantics, which specify the structure of sentences and
how they relate to facts in the world.

The interpretation of a sentence is the fact to which it refers.

If it refers to a fact that is part of the world, then it is true.

Inference is the process of deriving new sentences from old

ones. We try to design sound inference processes that
derive true conclusions given true premises. An inference
process is complete if it can derive all true conclusions
from a set of premises.

Alexander K. Seewald 35
alex@seewald.at / alex.seewald.at
 Knowledge and Reasoning

A sentence that is true in all worlds under all interpretations is

valid. If an implication sentence can be shown to be valid,
then we can derive ist consequent if we know ist premise.
The ability to show validity independent of meaning is
essential.

Different logics make different commitments about what the

world is made of and what kinds of belief we can have
regarding facts. Logics are useful for commitments they do
not make, because the lack of commitment gives the
knowledge base writer more freedom.

Alexander K. Seewald 36
alex@seewald.at / alex.seewald.at
 Knowledge and Reasoning

Propositional logic commits only to the existence of facts

that may or may not be the case in the world being
represented. It has a simple syntax and semantics.

First-order logic commits to the existence of objects and

relations in the world. It is useful for complex concepts.

Knowledge about actions and their effects can be represented

via a situation calculus. This knowledge enables the agent
to keep track of the world and to deduce the effects of
plans of action.

Alexander K. Seewald 37
alex@seewald.at / alex.seewald.at
 Knowledge and Reasoning
Knowledge engineering is concerned with building a useful
knowledge base. Knowledge acquisition is the process by
which the knowledge engineer becomes educated about the
domain and elicits the required knowledge.

The process of representing knowledge consists of deciding

what kinds of objects and relations (= the ontology) need
to be represented. Then a vocabulary is selected, and used
to encode general knowledge of the domain.

After encoding specific problem instances, automated

reasoning procedures can solve them - via a process
strongly related to search with admissible heuristics.

Alexander K. Seewald 38
alex@seewald.at / alex.seewald.at
 Knowledge and Reasoning

Good representations eliminate irrelevant detail, capture

relevant distinctions, and express knowledge at the most
general level possible, without being overly comprehensive

Constructing knowledge-based systems has advantages over

programming, but is not feasible for all problems.
Modeling relevant knowledge for a task may be infeasible.

State-of-the-Art are embedded AI systems, where AI is used

complementary to other programming techniques.

Alexander K. Seewald 39
alex@seewald.at / alex.seewald.at
Example: VIE-PNN
• Knowledgebased
system for neo-
natal nutrition
• Rules derived
from expert
knowledge.
• HTML-based
interface.
• In clinical use
for >5 years at
AKH Vienna

Alexander K. Seewald 40
alex@seewald.at / alex.seewald.at
 Planning

Planning agents look ahead to come up with actions that will

contribute to goal achievement. They differ from problem-
solving agents in their use of more flexible representations
of state, actions, goals, and plans.
Planning systems can be seen as efficient special-purpose
reasoning systems designed to reason about actions; or as
efficient search algorithms for the space of possible plans.

The STRIPS language describes actions in terms of their

preconditions and effects. It captures much of the
expressive power of situation calculus. Not all domains
and problems can be described in STRIPS.

Alexander K. Seewald 41
alex@seewald.at / alex.seewald.at
 Planning
STRIPS is too restricted for complex, realistic domains, but
can be extended in several ways; extensions of STRIPs are
still used in many realistic planning domains.

Hierarchical decomposition allows nonprimitive operators

to be included in plans, with a known decomposition into
move primitive steps. This is most effective when it serves
to prune the search space.

Many actions consume resources. It makes sense to treat

these as numeric measures in a pool. Time is one of the
most important resources. With a few exceptions, time can
be handled like any other resource.

Alexander K. Seewald 42
alex@seewald.at / alex.seewald.at
 Planning

It is not feasible to search through the space of situations in

complex domains. Instead we search through the space of
plans. For problems in which most subplans do not
interfere with each other, this will usually be efficient;
otherwise more complex domain-specific search strategies
are needed.

The principle of least commitment states that a planner

should avoid making decisions until they are needed.
Partial ordering constraints and uninstantiated variables
allows to follow a least commitment approach.

Alexander K. Seewald 43
alex@seewald.at / alex.seewald.at
 Planning
Execution monitoring is essential to ensure robustness.
Conditional planning takes failures into account when
planning; Replanning recomputes the whole plan on
failure. These are two points on a continuous spectrum.

Scheduling takes a given plan and creates an appropriate

schedule of execution. Scheduling can be formulated as
constraint-satisfaction problem, with time being treated
mostly like any other resource.

Automatic planners and schedulers have proven capable of

handling complex domains such as spacecraft missions and
manufacturing.

Alexander K. Seewald 44
alex@seewald.at / alex.seewald.at
Example: Shakey

Alexander K. Seewald 45
alex@seewald.at / alex.seewald.at
 Acting under Uncertainty
Uncertainty is inescapable in complex, dynamic or
inaccessible worlds; and means that many simplifications
that are possible with deductive inference are no longer
valid. Probability theory provides a way of summarizing
the uncertainty that comes from laziness and ignorance.

Basic probability statements include prior probabilities and

conditional probabilities over simple and complex
propositions. The joint probability distribution specifies
the probability for assigning values on all variables.

Bayes' Rule allows unknown probabilities to be computed

from known, stable ones.

Alexander K. Seewald 46
alex@seewald.at / alex.seewald.at
 Acting under Uncertainty
Conditional independence information is a vital and robust
way to structure information about uncertain domains.

Belief networks are a natural way to represent conditional

independence information. The links between nodes
represent the qualitative aspects of the domain, and the
conditional probability tables represent the quantitative
aspects.

The complexity of belief network inference depends on the

network structure. Inference mechanisms are of
exponential complexity in the worst case; in real domains,
the local structure makes inference more feasible.

Alexander K. Seewald 47
alex@seewald.at / alex.seewald.at
Example: Burglar alarm

Alexander K. Seewald 48
alex@seewald.at / alex.seewald.at
 Decision Theory
Simple decision problems can be solved by decision theory,
which relates what an agent wants (utility theory) to what
an agent should believe on the basis of evidence
(probability theory) Utility theory associates a utility
value to each state of the agent.
We can use decision theory to build a system that make
decisions by considering all possible actions and choosing
the one that leads to the best expected outcome. Such a
system is known as a rational agent.

Decision theory is normative - it describes rational

behaviour. It is probably not descriptive - people
systematically violate the axioms of utility theory.

Alexander K. Seewald 49
alex@seewald.at / alex.seewald.at
 Decision Theory

More complex sequential decision problems in uncertain

environments can be solved by calculating a policy that
associates an optimal decision with every state that the
agent might reach.

Methods to calculate optimal policies are closely related to

the general computational technique of dynamic
programming, which considers all possible paths in an
efficient way.

Alexander K. Seewald 50
alex@seewald.at / alex.seewald.at
 Question to the audience
What would you prefer?
A) 80% chance of winning 4000
B) 100% chance of winning 3000
[Allais, 1953] found that people strongly prefer B)

C) 20% chance of winning 4000

D) 25% chance of winning 3000
[Allais, 1953] found that people strongly prefer C)

No consistent utility theory for humans is possible!

0.8U( 4000)<U( 3000) and 0.25U( 3000)<0.2U( 4000)
cannot both be satisfied.

Alexander K. Seewald 51
alex@seewald.at / alex.seewald.at
 Communication
Agents need to communicate to each other and to the users.
Communication between learning agents is an active
research area which sheds light on the development of
language in humans.

Natural language processing techniques make it practical to

develop programs that make queries to a database, extract
information from texts, translate languages, or recognize
spoken words.

In all these areas, there exist programs that are useful, but
there are no programs that do a throrough job in an open-
ended domain.

Alexander K. Seewald 52
alex@seewald.at / alex.seewald.at
Shazam Entertainment

Alexander K. Seewald 53
alex@seewald.at / alex.seewald.at
 Agents as programming metaphor
• Procedural (classic) programming
• Declarative programming
• Object-oriented programming
• Constraint logic programming
• Event-oriented programming
• Knowledge-based software engineering
• Agent-based software engineering
...
Each of these gives an unique viewpoint on programming;
makes solving some problems easier and others harder. But
you still need a programmer!
For learning systems, you don't need a programmer. Most
of the work is done by learning systems.

Alexander K. Seewald 54
alex@seewald.at / alex.seewald.at
 Learning
Learning in intelligent agents is essential for dealing with
unknown environments; and for building agents without
prohibitive amount of work. All learning suffers from the
credit assignment problem = which steps are responsible
for a good or bad outcome?

Reinforcement learning is an active research topic, and

computationally very expensive. Temporal difference
learning and Q-Learning are common learning algorithms.

Genetic algorithms achieve reinforcement by increasing the

proportion of successful functions. They achieve
generalization by mutating and cross-breeding programs.

Alexander K. Seewald 55
alex@seewald.at / alex.seewald.at
 Learning

Learning a function from examples of its inputs and outputs

is called inductive learning. Learning in the inductive
setting is supervised and needs a set of training inputs and
outputs.

Unsupervised learning uses the structure of training data to

infer hidden relationships, which are harder to validate.

Inductive logic programming can learn relational

knowledge, as used in knowledge-based systems. This
kind of learning is generally very hard for larger problems.

Alexander K. Seewald 56
alex@seewald.at / alex.seewald.at
 Bias
"Bias refers to any criterion for choosing one generalization
over another other than strict consistency with the
observed training instances" (Mitchell, 1980)

Each learning algorithm is biased twofold:

• language bias = restricts possible concepts to be learned
• search bias = prefers certain models over others

Overfitting occurs when the structure of training data is

learned too well; and the generalization performance on
unseen data suffers.
Bias is essential to learning!

Alexander K. Seewald 57
alex@seewald.at / alex.seewald.at
 Learning
A large variety of learning algorithms is
available, which can learn:
• A state evaluation function to play checkers
• A belief network to model sleep stages
• A function to predict steel quality in production
• A function to predict insurance risks
• Logic programs to determine cancerogenity
• Association rules in supermarket basket analysis
• Time-dependent models of speech
• Response models of mailable customers
...

Alexander K. Seewald 58
alex@seewald.at / alex.seewald.at
 But learning is still hard! Why?
Inductive learning is inherently risky
• There is no safe way to predict the future.
• Bias is essential, but may be wrongly chosen.

No Free Lunch!
• Theoretically, it is not possible to learn anything.
• Practically, the world shows an enormous variety
of patterns. Life has adapted over billions of
years to take advantage of these specific patterns.

Alexander K. Seewald 59
alex@seewald.at / alex.seewald.at
 Example problem: Response Model
Problem: Not enough capacity to mail all customers.
⇒ Improve effectiveness by learning a response
model: 30% higher volume with same cost.

Alexander K. Seewald 60
alex@seewald.at / alex.seewald.at
 Example problem: Churn

Given: A set of customers with state, area

code, telephone number, and time/cost
information for calls in one month; plus
churn = have they switched to another
provider by the end of the month?

Create a useful model of customer churn,

so it can be reduced significantly!

Alexander K. Seewald 61
alex@seewald.at / alex.seewald.at
 Rules for Churn (1)
(total_day_minutes >= 245) and (total_eve_minutes >= 225.2) and
(voice_mail_plan = no) and (total_night_minutes >= 170.6) =>
churn=True. (64.0/0.0)
(total_day_minutes >= 236.9) and (total_night_minutes >= 230.6) and
(voice_mail_plan = no) and (total_eve_minutes >= 197.7) =>
churn=True. (12.0/1.0)
(total_day_minutes >= 223.3) and (total_day_minutes >= 264.8) and
(voice_mail_plan = no) and (total_eve_minutes >= 188) and
(total_night_minutes >= 132.9) => churn=True. (52.0/1.0)
(total_day_minutes >= 222.3) and (total_day_minutes >= 286.2) and
(voice_mail_plan = no) and (total_eve_minutes >= 150.8) =>
churn=True. (17.0/2.0)
(total_day_minutes >= 221.9) and (total_eve_minutes >= 261.6) and
(voice_mail_plan = no) => churn=True. (41.0/7.0)

Alexander K. Seewald 62
alex@seewald.at / alex.seewald.at
 Rules for Churn (2)
(number_customer_service_calls >= 4) and
(total_day_minutes <= 160) and (total_eve_minutes <=
233.2) and (total_night_minutes <= 254.9) =>
churn=True. (69.0/0.0)
(number_customer_service_calls >= 4) and
(total_day_minutes <= 182.1) and (total_eve_minutes
<= 190.7) and (total_night_minutes <= 285) =>
churn=True. (22.0/0.0)
(number_customer_service_calls >= 4) and
(total_day_minutes <= 135.9) and (account_length >=
72) => churn=True. (14.0/0.0)
(number_customer_service_calls >= 4) and
(total_eve_minutes <= 135) => churn=True. (12.0/4.0)

Alexander K. Seewald 63
alex@seewald.at / alex.seewald.at
 Rules for Churn (3)
(international_plan = yes) and (total_intl_minutes >=
13.2) => churn=True. (54.0/0.0)

(international_plan = yes) and (total_intl_calls <= 2) =>

churn=True. (50.0/0.0)

=> churn=False. (2926.0/91.0)

Alexander K. Seewald 64
alex@seewald.at / alex.seewald.at
 Demo

Demonstration of the WEKA Machine

Learning Workbench

Open Source, available at

http://www.cs.waikato.ac.nz/~ml/weka

Integrated into Pentaho' s Open Source

Business Intelligence Suite
http://www.pentaho.com

Alexander K. Seewald 65
alex@seewald.at / alex.seewald.at
 Past Projects
2000-2005 Employed at OFAI as junior researcher
2001 EEG data analysis (contributed by Brain Research institute, Vienna)
2000-2002 A New Modular Architecture for Data Mining (FWF)
2002 3DSearch (multi-document summarization, EU & uma AG)
2002-2003 Intelligent Go Board (embedded device to capture moves of
Japanese Go during play, presented at Innovation Workship in ' 05)
2003-2005 BioMinT (integrated system for biological text mining, EU FP5)
2004-2006 SA Train (Spam training methodology for SpamAssassin, Evaluation
of commercial and open-source spam filter systems)
2005 Digits (handwritten digit recognition: open source corpus and
preliminary experiments)
2006 Employed at GE Money Bank as CRM Analyst
2006-2007 IGO-2 (image mining on images of Go final board states)
2007 Websuit (image mining on GFP/DIC images contributed by Univ. of
Colorado at Boulder; related to my recent ERC Ideas proposal)
2007- Employed at Ikarus in R&D for spam filtering and virus detection

Alexander K. Seewald 66
alex@seewald.at / alex.seewald.at