Title: Artificial Neural Networks ANNs Computing Technology and Applications by Sdhabhon Bhokha Sdhabhonait
1Artificial Neural Networks (ANNs)Computing
Technology and ApplicationsbySdhabhon Bhokha
(Sdhabhon_at_ait.ac.th http//www.nyakobo.com/)28
June 2000
- ContentIntroduction Artificial Intelligence
(AI) 2-12Artificial Neural Networks
(ANNs) 13-28Illustrative Applications 29-42
2Human Intellectual (How does a man become more
clever?)
3(No Transcript)
4(No Transcript)
5Computer or Hardware
- Von Neumann Computer (Serial Processor)
- It is not definitely compatible with ANNs
concept - However, it can be use in ANNs computing (as it
is still being used today) - Parallel Processors
- They are fully support the ANNs concept and work
but only in research work
6Generalization
- Capability to reasoning or solving even the
problem which are beyond the learning (or the
knowledge being). - Generalization makes the AI different from and
have edged over all the existing rule-based
systems. - Generalization is an ability to respond to
patterns which are not presented during learning
or ability to generate unknown situations from
the known experience or examples, then can derive
the output for unfamiliar inputs
7(No Transcript)
8(No Transcript)
9Characterization between Neural Network
versusRule-based System (source Klemic, 1993)
Rule-based system
ANNs
Ruleless
Rule-based
Distributed
Symbolic
Parallel
Serial
Discrete
Continuous
Problem-solving
Pattern recognition
Neurophysiology
Psychology
Cognitive
Behavioral
Structural
Functional
Logical
Intuitive
10Characterization between Neural Network
versusRule-based System (source Klemic, 1993)
ANNs
Rule-based system
Example based
Rule based
Domain free
Domain specific
Finds rules
Needs rules
Little programming needed
Much programming needed
Little programming needed
Difficult to maintain
Not fault tolerant
Easy to maintain
Need (only) a database
Need human expert
Need (only) a database
Rigid Logic
Adaptive system
Requires re-programming
11(No Transcript)
12When did ANN start? and How did ANNs start?
13Five Decades in AI history (Forsyth, 1992)
- 1950s Dark Ages (Neural Networks)
- 1960s The Age of Reason (Automated Logic)
- 1970s The Romantic Movement (Knowledge
Engineering) - 1980s The Enlightenment (Machine Learning)
- 1990s Gothic Revival (Neural Nets Revisited!)
14Conventional Steps of ANNs Modeling
- 1. Data acquisition (only for the supervised
learning) - 2. Problem representation (independent/dependent
variables and forms, i.e. real or binary) - 3. Architecture and paradigm (layers, nodes,
links, and learning algorithm) - 4. Training the network
- 5. Testing for validation
- 6. Application (and routine maintenance)
15(Expected) Advantages of ANNs Compared to
Conventional Programming (Rule Based System)
- Simple calculation
- Generalization capability
- Fault tolerance
- Fast processing (In application mode)
- Dynamic
- Distribution free
16Applications of ANNs in Civil Engineering
- Diagnosis,
- Pattern Recognition or Pattern Interpretation,
- Forecasting (Predicting),
- Decision Making,
- Optimization
- Fault Detection
17Computer Program or ANNs Software
- Commercial software package (various option,
architecture, paradigms but expensive and not
flexible for modification) - Written for specific objectives and functions
e.g. spreadsheet program, Basic, C, C, PASCAL,
FORTRAN
18Awareness of Using ANNs
- Lack of friendliness (No explanation provided),
- Thus, the user needs two background or knowledge
in 1)problem domain 2) ANNs and 3) computer
technology . - Sometimes, developing and maintaining the AI may
Time-consuming and expensive process, which
will be the disadvantage. - There is no guarantee of success on achievement.
19How Neuro-cell works
Dendrite
Axon
Dendrite
Neuro-cell 1
Neuro-cell 2
201
U11
Omk
Tmk
V11
U21
j 1
k 1
2
f(Nbj)
Nbj
U31
i 3
Nbk
f(Nbk)
f(Nb.)
1
0.5
Nb
0
Sigmoid Transformation Function
21Other Names of ANNS
- ANN s topology or architecture likes humans
brain, i.e. consists of neuro-cell (or neuron) - Other names of ANNs are Connectionist Model,
Parallel Distributed Processing Model,
Neuromorphic System, Neural Computing
22Data Acquisition
- Necessary for problem representation and
modeling, i.e. the inputs, output, topology and
learning algorithm - Sources of data are 1) experts questionnaire
(same as KBES does) 2) historical records and
3) simulation. - What should the samples have?
- Adequacy (depended on the parameters to be
estimated) - Completeness They can represent characteristics
and distribution of variables, includes the
extreme cases of the samples (if possible)
23Variables
- Independent variables or inputs (real or binary
and they can be mixed !) - Dependent variables or outputs (real or binary
(depended on the transformation function) - Remark Binary representation use 0 or 1 to
represents Yes or No, On or Off, Black
or White, Good or Bad, Easy or Complex
24ANNs Programming
- Node (or Neuron, Processing Element, Processing
Unit) acts as a neuro-cell - Link is the connection between the nodes where
the input or output can flow through . Two kinds
of links are 1) fully and 2) patterned link. - Weight on each connectionist a number (instead of
hormone or Chemical-Electrical reaction in
between the axon of a cell and dendrite of the
other)
25Hidden Layer (s), Hidden Node(s), and Pruning
- More hidden layer, and nodes tend to increase
generalization capability of the network but
learning process may need more samples and
training time - Appropriate No. of hidden layer and node can be
determined by constructive or destructive (or
pruning) approaches. It is the process by which
the redundant units can be eliminated from the
trained larger network to produce the smaller
network that can perform the same task (at same
capability) as the larger network - Not only the hidden layer and node, but also the
input and output nodes can also be pruned
26Transformation Function
- Transformation function (which sums up the
product of inputs and weights, and fire or
excitate the outputs) Various forms of
transformation function, e.g. sine, hyperbolic
tangent, logistic (Sigmoidal, hard limiter) They
must have at least 2 important characteristics
1) bounded and 2) differentiable.
27ANNs Learning Process
- Learning process consists of training and
testing. The ANNs can learn by means of
adjusting the weights (At the beginning, the
weights have to be initiated, i.e. by random
process). Then, learning process becomes an
iterative fashion - Supervised learning (needs adequate samples for
both training and testing), e.g. Back-propagation
net and Generalized Delta Rule (BP - GDR) - Unsupervised learning (does not need the
samples), sometimes called Self learning),
e.g. Adaptive
28Learning Rate and Momentum,
- Learning rate (0-1), in practice use 0.6 to 0.9
(steps by 0.1) - Momentum (0-1), steps by 0.1 (to minimize the
parameters, null momentum is preferred) - High learning rate and momentum tend to converge
rapidly but sometimes risk to fall into a local
minima, and vice versa
29Criteria to Stop Training
- Nos. of Epoch ( 1 Epoch consists of n cycles
where n is numbers of training samples) - Sum Squared Error (SSE) or Mean Squared Error
(MSE)
30Function of building
Structural system
Construction Cost Duration
Height
Site accessibility
Complexity of foundation
Exterior finishing
Major Factors Affecting Construction Cost and
Duration of Buildings
31(No Transcript)
32Monthly Price Indices for Three Major
Construction Materials during 1987 - 1997
33Buildings Classified by the Year Start
Constructing (136 projects)
34Statistics for 136 Building Samples
- 1. Building functions (2-binaries) 63
residences (46.3), 59 offices (43.4), 6 dual
functions (4.4) and 8 others (5.9). - 2. Structural systems (2-binaries) 29
cast-in-situ RC (21.3), 100 post-tensioned
slabRC frame(73.5), and 7 others (5.1) - 3. Height (1-binary) 69 normal height (50.7)
and 67 extra height (49.3). - 4. Foundation index (1-binary) 101 complex
(73.4), 35 simple (25.7). Examples for complex
foundations are retaining and diaphragm walls,
and mat footing.
35- 5. Exterior finishing (2-binaries) 58
brick/cement block wall (42.7), 77 frameglass
or cladding (56.6), and 1 prefabricated wall
(0.7). - 6. Interior decorating (1-binary) 88 normal
(64.7), and 48 excellent (35.3). Excellent
interior decorating consists of premium class of
1) sanitary fixture, accessories 2) luxury
electrical accessories 3) carpet, floor tile,
and wall paper and 4) built-in furniture. - 7. Site accessibility(1-binary) 43 poor
(31.6), and 93 good (68.4). - 8. Price indices (2-real values) average price
indices for steel and cement are 131 and 126,
respectively,
3620,000
Variations in Construction Cost of high-rise
buildings in Greater Bangkok Area during 1987 -
1995 (136 projects)
37Variation in Construction Duration of High-rise
Buildings in the Greater Bangkok Area during 1987
- 1995 (136 projects)0
38Description of the Successfully Trained Networks
- Cost Duration
- Input, hidden, output nodes 12, 6, 3 11, 6, 1
- Learning rate (momentum) 0.6 (0) 0.6 (0)
- Training time, sec 1,336 1,885
- Epoch 23,235 44,700
- SSE (training set) 6.28x10-4 7.97x10-6
- SSE (test set) 1.40x10-3 2.11x10-4
- MSE (training set) 0.0030 0.0003
- MSE (test set) 0.0045 0.0018
39Accuracy of Networks Forecasting (136 samples)
- Cost network
Duration network - Underestimated -46.5 to -0.2 0.1 to 136.8
- Nos.of project 58 (42.7) 40 (29.4)
- Over estimated 2.4 to 150.0 2.4 to 150
- Nos. of project 78 (57.3) 96 (70.6)
- Average errors
- underestimated -4.8 -4.6
- overestimated 9.2 18.2
- overall 4.4 13.6
40Variation of Forecasting Errors by 12, 6, 3
Network in according to Construction Costs (136
samples)
41Variation of Forecasting Errors by 11, 6, 1
Network in according to Construction Duration
(136 samples)
42Shift of Paradigms (New Fashions)
- Combined ANNs-KBES, i.e. the KBES will be used
for explanation function - Combined ANNs-GA, i.e. in analysis, design and
optimization, the GA will function as learning
rule - Combined ANNs-DBMS, i.e. uses DBMS for storing
the train or test data, routine maintenance