|
Areas for Exam 1 Questions (10 questions will be
selected):
-
Be prepared to
summarize the basic advantages and disadvantages of: fuzzy vs neural network
systems, conventional vs neural network signal processing, and/or
conventional vs fuzzy expert systems.
-
Be prepared to to
set up a linear and nonlinear dynamic system using state and output
equations (identifying states, inputs, and outputs).
-
Systems
terminology: know the difference between linear vs nonlinear, dynamic vs
static/quasi-static, and/or deterministic vs stochastic.
-
Controls
Terminology: Be prepared to state the advantages-disadvantages of
feedback vs feedforward neurocontrol, linear vs adaptive control, and/or
linear vs fuzzy control.
-
Know the following
"fuzzy" terminology: fuzzy sets (e.g., vs crisp, universe of
discourse, membership function, basic properties and math), fuzzy relations
and inference.
-
Be able to
describe the Mamdani and Suguni fuzzy inference system structures, and how
they differ.
-
Be able to
describe how you would set up a Mamdani system, using GFIE, given linguistic
problem description.
-
Be able to
graphically go through the process of Mamdani inference for a specific
expert system example, from inputs to output.
-
Be able to
identify and give mathematically 2 approaches for "and" 'or"
"implication" aggregation" and "defuzzify" (this
includes solving for a centroid).
-
Be able to
describe how you would set up a fuzzy control system, and why such a system
may have some advantages over a conventional control system. Give an
example.
-
Be able to
describe, with a concrete example, how changes in an internal parameter
(e.g., membership function parameter) or the assumed type of operation
(e.g., "and") affected fuzzy system performance.
-
Be able to map
between a "rule table" to linguistic description of rules.
-
Be able to
describe the basic anatomy and physiology of a neuron, basic neuromuscular
terms (Table E), structural organization of the brain (Table F), and/or
basic pathology of key diseases (Table J). Expect several short parts
to this question.
Areas for Exam II Questions (10 will be selected, 9 of which are from the
following list):
-
Mathematically
describe the classic artificial "neuron" processing element,
including both the input "integration" function and the activation
(transfer) function (give for a "solf saturating" sigmoid).
Given three inputs (on <0,1>, determine the output.
-
The
classic unsupervised approach for ANN learning is Hebbian. Describe
and give the mathematics for the special variant of a "forgetting"
term, then one other of your choice. How does it differ from the Delta
rule?
-
Given
a feedforward net, give and be able to apply the classic backprop learning
algorithm for an internal neuron (both the forward (action) net and backprop
(e.g., see Fig 10.8)
-
Set
up a neural net structure that can be used to solve a small Sugino fuzzy
inference system where some of the ANNs use radial basis function
"receptive fields." (e.g., see Fig 9.10). Give the
conditions under which an RBFN & FIS are functionally equivalent (e.g.,
see p. 242).
-
Draw
a block diagram structure for a neural network model reference controller
and plant, and then describe how you would go about setting up plant
identification and controller training using Matlab (e.g., what parameters
are needed for the controller training algorithm?)
-
What
is reinforcement learning, and why is implementation so challenging?
Carefully delineate it from supervised learning, and then describe several
of the key approaches.
-
Draw
the basic Cerebellar circuit, and using it, discuss how it might learn to
recognize patterns (e.g., ideas of Marr, Albus, etc).
-
One
of the three topics we discussed related to Houk et al.
-
What
are the classic steps for setting up and using a genetic algorithm?
-
Describe
the protocol for gradient-free optimization, and then describe a structure
for applying it to the use of GA's for neuro-fuzzy control.
|