Cse 481 Final Word List

Front 1

a priori

Back 1

derived by logic, without observed facts.

Front 2

abductive reasoning

Back 2

if p then q………q therefore p is not a valid deductive principle. This is called abduction.

Front 3

activation level

Back 3

based on the cumulative strength of its input signals. Each input signal is based on the connection weight.
The sum of the weighted inputs.

Front 4

analogical reasoning

Back 4

Assumes that if two situations are similar in some respects they will be similar in others (pg 409)

Front 5

associationist

Back 5

Believe that we define the meaning of an object in terms of a network of associations with other objects. This can be displayed using a semantic network.

Front 6

associative memory

Back 6

Works by recalling information in response to an information cue. Type of memory that when searched returns a particular value. There are three different types:
auto-associative just returns X if it exists
heter-associative returns the Y corresponding to the closes Xi to the X you typed in
interpolative takes your X, it transforms your X into the closest Xi. Then it then looks at the Yi corresponding to the Xi and applies a transformation that’s analogous to the transformation from X to Xi and returns this newly transformed Yi.

Front 7

atomic actions

Back 7

Actions that cannot be broken down in to smaller actions. For example, in robotics, turn the motor one revolution.

Front 8

attractor networks

Back 8

networks that employ feedback connections to repeatedly cycle a signal within the network. The network output is considered to be the network state upon reaching equilibrium.
an attractor is a state in the network which other states, that lie within its basin, evolve to over time.
these type of networks are also known as “memories”

Front 9

autonomous

Back 9

a system that can interact with its environment without the direct intervention of any other agents. Thus, It must have control over its own actions and internal state.
another type of agent-oriented problem solver.

Front 10

backpropagation

Back 10

a way of placing blame on intermediate steps by splitting up the error based how much significance they had in creating the answer.
A multilayer feedforward neural net architecture which uses the supervised mode of learning. This is the most widely used type of neural net.

Front 11

bagging

Back 11

takes some samples from the training instances (and then put them back). When sample, we replicate some of these training sets.
This helps us to produce multiple classifiers which will form into a composite classifier by each one voting (all have same vote strength).

Front 12

BAM

Back 12

described by Bart Kosko consists of two fully connected layers of processing elements. The vectors on a are taken from the set of Hamming vectors. network can be transformed into an autoassociative network by using the same weight initialization formula on the set of associations. is used to retrieve patterns from memory by initializing the X layer with an input pattern. (pg 496)

Front 13

Bayesian belief network

Back 13

focuses only on the things that are relevant. However there is a danger of eliminating something relevant accidentally. Thus you build an inductive bias into your program with these assumptions. Also uses certainty measures.
Offers a computational model for reasoning to the best explanation of a set of data in the context of the expected causal relationships of a problem domain.

Front 14

boosting

Back 14

we look at all of our training instances at each replication. When looking we assign a weight to each instance that is in the training set. This weight corresponds with that particular set’s (vector’s) importance. As we adjust these weights our different classifiers are produced.
The reason this happens is because the learner will focus on whatever has the most weights. When we change these weights to different instances, the learner looks at different instances and thus creates new classifiers.
This clearly helps us to produce multiple classifiers which will form into a composite classifier by each one voting (the more accurate the component classifiers the stronger their vote).

Front 15

candidate elimination

Back 15

ombines two algorithms to reduce the version space (a set of all concept descriptions that go along with the training set). Maintains two sets of candidates: maximally general(G) and maximally specific(S). Then, the algorithm specializes G and generalizes S until they both converge on a single candidate. The combination turns into a bi-directional search.
This has several advantages such as: eliminating the need to save instances and DOES NOT requires all training examples to be present before it starts “learning”.
This is a type of supervised learning.

Front 16

canonical form

Back 16

When working with conceptual graphs, we can follow these rules that have a “subtle but important property of preserving meaningfulness”. This kind of stuff is important when using conceptual graphs to implement natural language understanding.
NOTE: basically introduces a rule that says “if two sentences have the same meaning, they will be graphed both syntactically and semantically identical.”

Front 17

case frame

Back 17

collecting primitives like agents, time, object and making them into a frame like structure. Then applying situations like “sally fixed her chair with glue” (page 235)

Front 18

case-based reasoning

Back 18

A problem-solving system that relies on stored representations of previously solved problems and their solutions. Also allows programs to “learn” from its own experiences.
Does not require extensive analysis of domain knowledge
You can fill empty cases by:
retrieving cases form memory
modify a similar one to the one you’re missing
make that the rule for the empty case

Front 19

cellular automaton

Back 19

families of simple finite-state machines that exhibit interesting, emergent, behaviors through their interactions in a population.
Think of an infinite, regular grid of cells, each in one of a finite number of states. Every cell has the same rule for updating, based on the values in this neighborhood. Each time the rules are applied to the whole grid a new generation is produced.
“Game of Life.”

Front 20

chunking

Back 20

refers to a short-term memory mechanism and techniques to exploit it.

Front 21

closed world assumption

Back 21

in traditional logic and programming, if you don’t explicitly say that it is true, then we make the “closed world assumption” that it is false. These follow monotonicity. You always add conclusions/knowledge never take any away. Once its in there, its in there.

Front 22

common super/sub type

Back 22

Being in the same set as something else. Kind of self explanatory.
Common sub: if s, t, and u are types and t < s and t < u. Then t is a common sub of both s and u.

Front 23

conceptual clustering

Back 23

trying to discover useful categories in unclassified data. You begin with a collection of unclassified objects and a way of measuring the similarity between each of the objects.
The goal is to organize the objects into classes that meet some standard of quality.
This is a machine based learning approach.

Front 24

conceptual graph

Back 24

finite, connected graph split into two parts. The nodes of the graph are either concepts or conceptual relations. Each graph represents a single proposition.
NOT use labeled arcs. The conceptual relations take the place of these.

Front 25

conceptual model

Back 25

when the knowledge engineer talks to the domain expert they construct a conceptual model in order to illustrate the domain knowledge, this helps to actually determine the construction of the knowledge base.
lies between human expertise and the implemented program

Front 26

connectionism

Back 26

belief that design does not need structured symbolic sentences.

Front 27

decision tree

Back 27

A representation that allows us to determine the classification of an object by testing its values for certain properties. (pg 408)

Front 28

crossover

Back 28

a way of mixing two chromosomes by splicing pieces of each one into the other one. Two parental programs are selected based on fitness.

Front 29

defeasible

Back 29

knowledge base that can be added and undone. Unlike closed world thinking.

Front 30

delta rule

Back 30

generalization of the perceptron learning algorithm that is used many neural networks, including back-propagation.
Instead of the hard-limiting function used in perceptrons, it uses a smoother, continuous function (also known as sigmoidal). A common sigmoidal activatin function is the “logistic function”.

Front 31

demon

Back 31

portion of a program that is not invoked explicitly, but that lies dormant waiting for some condition(s) to occur.
Invoked as a side effect by some other action in the knowledge base

Front 32

Dempster-Shafer theory of evidence

Back 32

Addresses the problem of measuring certainty by making a distinction between the lack of certainty and ignorance.
Great for when it is important to make a decision based on the amount of evidence that has been collected.

Front 33

Distributed

Back 33

a type of representation of intelligence. Used in agent-based computing.

Front 34

domain knowledge

Back 34

added by the “system builder”

Front 35

dynamic programming

Back 35

a type of reinforcement learning inference algorithm that computes value functions by backing up values from successor states to predecessor states. The methods systematically update one state after another, based on a model of the next state distribution. (pg 447)

Front 36

evolutionary learning

Back 36

learning based around methods mapped and inspired from evolution.

Front 37

expert system

Back 37

A type of application program that makes decisions or solves problems in a particular field by using knowledge and analytical rules defined by experts in the field.

Front 38

explanation-based learning

Back 38

uses an explicitly represented domain theory to construct an explanation of a training example, usually a proof that the example logically follows from the theory. By generalizing from the explanation of the instance, rather than from the intance itself, it filters noise, selects relevant aspects of experience, and organizes training data into a systematic and coherent structure.(pg 424)

Front 39

family resemblance theory

Back 39

by Wittgenstein argues that categories are defined based on the relationship and similarities between the members of the group, NOT by some necessary and sufficient conditions.

Front 40

fitness function

Back 40

deciding to keep the top certain percentage of generated chromosomes (or something else). For example, only the top 70%, the most fit, make it to the next round.

Front 41

frame

Back 41

a static data structure used to represent stereotype situations
A formal method of representing information about a single idea or concept in terms of properties where the information is stored in slots.
Extends semantic networks by making it easier to organize our knowledge hierarchically and allowing complex objects to be represented as a single frame rather than a large network structure

Front 42

frame problem

Back 42

the problem of representing the side effects of actions.

Front 43

fuzzy logic

Back 43

logic that consists of not just 0’s and 1’s but also .5’s and ‘3’s. For situations like “matt is tall”. Can’t really say yes or no, but give a more relative answer

Front 44

genetic algorithm

Back 44

An algorithm that mimics evolution and natural selection to solve a problem by creating chromosomes based on a few decided strategies (probably conditional statements). Then setting up these chromosomes against each and mutating them until you achieve optimal values for each of the strategies.

Front 45

goal regression

Back 45

matches the generalization goal with the root of the proof tree, replacing constants with variables as required for the match. (pg 427)

Front 46

goal-directed preference

Back 46

Organizing cases can be difficult. GDP helps us by organizing cases by goal descriptions. Retrieves cases that have the same goal as the current situation.

Front 47

Hebbian learning

Back 47

as your brain does the same task over and over again, a “rut” gets created and completes that same task quicker. Based on observations in biology when one neuron contributes to the firing of another neuron, the connection or pathway between the two neurons is strengthened. (pg 484)

Front 48

Homomorphic

Back 48

implies a one-to-one correspondence between objects and actions in the world and the computational objects and operations of the programming language

Front 49

Hopfield nets

Back 49

a special type of attractor network who’s convergence (ending output) can be represented by energy minimization.
Can be used to solve constraint satisfaction problems (like the traveling salesperson problem) by mapping optimization function to energy function (pg 500)

Front 50

hybrid design

Back 50

Combining different reasoning models, for example case based and rule based reasoning. It is good because integrated paradigms get a cooperative effect where the strengths of one system compensate for the weakness of another.

Front 51

Induction

Back 51

using previous knowledge and given facts and making your “best guess” about a conclusion.

Front 52

inductive bias

Back 52

choosing to ask particular questions by making assumption and disregarding certain ideas. Refers to any criteria a learner uses to constrain the concept space.

Front 53

inference engine

Back 53

The processing portion of an expert system. With information from the knowledge-base, the inference engine provides the reasoning ability that derive inferences (conclusions) on which the expert system acts. A small part of the program that interprets the rules to be used.

Front 54

knowledge base

Back 54

the heart of the expert system. Contains all of the knowledge (usually if.. then statements) of the application domain.
Contains both general knowledge and case-specific knowledge

Front 55

knowledge engineer

Back 55

Their main task is to select the software and hardware tools for the project and help the domain expert transform their knowledge into the knowledge base.

Front 56

law of the excluded middle

Back 56

States an element can not belong to both a set and its complement.

Front 57

marker

Back 57

a unique token contained in a conceptual graph. Instead of a node like “dog” it’s a node like “dog:#668584” which is a specific dog with specific attributes.

Front 58

Markov model

Back 58

A graph that is directed where the probability of arriving at any state S1 from the set of states S at a discrete time T is a function of the probability distribution of its being in previous states of S.
Each state si of S corresponds to a physically observable situation.
It seems like the word “observable” goes along with “Markov Model.”

Front 59

medium

Back 59

programming languages are the medium whereas data structures are the scheme.

Front 60

minimum distance classification

Back 60

A discriminant function evaluates class membership (or importance) based on the distance from some central point. Classification based on this discriminant function is called minimum distance classification.
Classes that are linearly separable can have a minimum distance classification.

Front 61

Monte Carlo method

Back 61

A type of reinforcement learning inference algorithm. The method does not require a complete model. It instead samples the entire trajectories of states to update the value function based on the episodes' final outcomes. It does not require experience , a sample sequences of states, actions and rewards on-line or simulated with the enviroment. It solves the reinforcement learning problem by averaging sample returns. To ensure well defined returns the methods are defined only for full episodes. (pg 448)

Front 62

mutation

Back 62

take a random number and apply it to one of the values in the chromosomes.

Front 63

negation as failure

Back 63

a way to prove something by proving that its negative cannot be true. This works because in certain languages like PROLOG anything we don’t explicitly state as true we assume to be false.

Front 64

network topology

Back 64

the pattern of connections between the individual neurons. This ends up being the most significant factor in determining a net’s inductive bias.

Front 65

neural network

Back 65

it uses neurons for problem solving. Its topology is its primary source of inductive bias.
Cannot represent an idea like “Matt is taller than Justin” according to connectionists.
A neural network is designed as an interconnected system of processing elements, each with a limited number of inputs and outputs. Rather than being programmed, these systems learn to recognize patterns.

Front 66

non/monotonic

Back 66

based on how things usually work. Similar to frames.

Front 67

opportunistic search

Back 67

whenever rules fire to conclude new information “control moves to consider” those rules which have that new information as a premise this makes any new concluded information the controlling force for finding the next rules to fire

Front 68

Perceptron

Back 68

inputs and outputs of +1 or -1 based on weights and threshold values. Learns based on training examples. Also takes in to consideration a learning rate.
Can only solve problems that are linearly separable (thus it can have a minimum distance classification).
Perceptrons minize the error of only the training set you give them. It only fits the data that you give it.
For better inputs, use a logicistic function

Front 69

planning

Back 69

Finding a sequence of actions that allows a problem solver to accomplish some specific task. Must note things in the environment that are changing as well as things that are not. (because we have to keep track of side effects in coincides with frame problem)
Where we organize explicit knowledge to control problem solving in complex domain. Examples: robotics.

Front 70

premise

Back 70

One of the propositions in a deductive argument. If X then Y, X is the premise.

Front 71

production rule

Back 71

conditional statements including high-level if and then statements. Something about “update working memory”.

Front 72

referent

Back 72

the individual that we are currently looking at in a conceptual graph. As Tiffy would say “the ID number associated with the concept node.” Examples: Dog; Emma (Emma is the referent) Dog; #123 (#123 is the referent)

Front 73

reinforcement learning

Back 73

places an “agent” in an environment and it receives feedback. In this case, learning requires the agent to act and then to interpret the feedback from those actions.
This is different from supervised learning because it does not have a “teacher” per say, but instead the agent must create a policy for interpreting all feedback.

Front 74

representation

Back 74

the best way to capture the significant parts of the intelligent activities we go through. Examples: language, use of motor skills.

Front 75

RETE

Back 75

an algorithm that helps optimize a search for rules in the knowledge base of an expert system by matching rules to the data by using a pointer to the rule. Greatly enhances the speed and execution of the system without destroying the semantic behavior.

Front 76

salient-feature preference

Back 76

Organizing cases can be difficult. This preference matches cases based on the most important features or matching the largest number of important features.

Front 77

scheme

Back 77

data structures are the scheme whereas programming languages are the medium.

Front 78

script

Back 78

takes a sequence of events that often happens on a daily basis and makes a structured representation.
Often used in natural language understanding systems
Broken up in to scenes

Front 79

semantic network

Back 79

A graph consisting of nodes that represent physical
or conceptual objects and arcs that describe the relationship
between the nodes, resulting in something like a data flow
diagram

Front 80

set cover

Back 80

There is a set cover approach to abduction. It defines an abductive explanation as a covering of predicates describing observations by predicates describing hypotheses.
In other words, the set cover approach to abduction attempts to explain the act of adopting a revocable belief in some explanatory hypothesis on the grounds that it explains an otherwise unexplainable set of facts.

Front 81

situated

Back 81

an agent that receives input from the environment that it is active in and can ALSO effect changes within that environment. Examples: internet, game playing.

Front 82

strong method

Back 82

a method of problem solving that involves knowledge specifically about the domain of the problem. Its explicitly encoded into the program.

Front 83

subsumption architecture

Back 83

Just a bunch of task-handling behaviors. Each behavior is a finite state machine that continually maps some perception based input into an action oriented output. Production rules drive this in a blind sort of way. Basically like a little lego mindstorms toy.
Brooks questioned the need for any centralized representational scheme and said that “I will attempt to show how intelligent beings might evolve from lower and supporting forms of intelligence.”
Intelligent behavior emerges from the interactions of simple architectures

Front 84

supervised learning

Back 84

Where the program basically has a teacher that is able to tell it whether an instance is a positive or negative example of a target concept.
Organization and training of a neural network by a combination of repeated presentation of patterns, such as alphanumeric characters, and required knowledge. An example of required knowledge is the ability to recognize the difference between two similar characters such as O and Q. Synonym: learning with a teacher. Contrast with self-organized system; unsupervised learning.

Front 85

syntax

Back 85

rules for combining words into legal phrases and sentences. Also can be involved with numbers or code.
Does not have meaning like semantics.

Front 86

threshold function

Back 86

produces an “on” or “off” state by computing the final output state of neuron based on a threshold value. If the neuron is to far above or below the threshold value, it is “off”, otherwise “on”.

Front 87

transformational analogy

Back 87

possibly something to do with interpolative associative memory.

Front 88

truth maintenance

Back 88

Representations and search procedures that keep track of conclusions that might later need to be questioned. In defeasible reasoning the TMS preserves the consistency of the knowledge base.

Front 89

weak method

Back 89

Examines only the syntactic form of states to try to come up with its solution. Intended to solve a wide variety of problems.

Front 90

winner-take-all

Back 90

selects the node whose pattern is most like the input vector and adjusts it to make it more like the input vector. It is unsupervised in that “winning” is simply seeing which node has a current weight vector closest to the input vector.
An algorithm that works with the single node in a layer of nodes that responds most strongly to the input pattern. The learning is unsupervised because the winner is determined by a “maximum activation” test. The weight vector of the winner is rewarded by bringing its components closer to those of the input vector. It can be viewed as a competition among a set of network nodes. (pg 474)