50 terms

Organic Computing

Organic Computing Lecture at KIT (www.kit.edu) - Learning Cards by Alexander Wolf
Name some Organic Computing Scenarios
Smart house
Smart car
Smart shop/warehouse
Smart clothes/wearable computers
Reasons for Organic Computing
- Future: large collections of interacting intelligent devices and networks
- Impossible to explicitly control them directly
- Need for self-organisation and life-like properties of these systems
Definition of organic computing
- Biologically-inspired computing
- Organic properties → life-like
- Learning, adaption, emergence, self-x-properties
5 Essential properties of OC
What is the difference between autonomic and organic systems
Autonomic Computing:
-focus on architecture,
- remove human user from system
Organic Computing:
- focus on collection of interacting devices
- allow human interaction for objectives
OC - Architecture requirements
o Support self-organization
o Allow external control
o Allow to observer, analyze and characterize the current internal/external system state
o allow future prediction
o support creation of new behaviour
o support learning
System under observation and Control - name the elements
Input, Output, Internal state s
O/C - Observer elements
O/C - Controller elements
Action Selector
History (action-> situation)
Adaption module
simulation module
How can OC Systems be realized (topology)
Central (adaptive, top-down)
Distributed (self-organizing, bottom-up)
Multi-Level (controlled self-organizing, up-down)
What types of control action do you know?
Control environment
Control communication
Control local behaviour of components
Outline MAPE
Monitor Analyze Plan and Execution
- for enterprise server architectures -
- Managed Elements contain
-- Autonomic element
-- Autonomic manager with:
---- Monitor
---- Analyser
---- Plan
---- Execution
---- Knowledge
What are the differences between MAPE and O/C-Architecture
- Knowledge kept in one unit
- sequence of operations instead of architectural components
- no objectives from user, but from policies
- no offline-learning
What is the difference between strong and weak self-organisation
self-organization --> system can re-organize itself
strong self-organization: no explicit central internal or external control
weak self-organization: system with internal central control
characteristics of a self-organising system
Regular input variables
Control input variables
Behaviour ß
Output variables
What is the hierarchy of notions of self-organisation
How can one calculate the degree of self-organization
complexity reduction R = Vint - Vext
static degree of self-organisation S=(Vint-Vext)/Vint
What is emergence
Local action in a self-organizing system leads to new global patterns -- "the whole is more than the sum of its parts"
How can emergence be calculated
Entropy H = -SUM(p_j * ld(p_j))
Hmax under equal distribution
Emergence e = Hmax-H
Name common self-x-properties
What are the characteristics of self-x-properties
Resource allocation
Coordination mechanisms
Digital Pheromone Coordination
Market based coordination
Tag based coordination
Token based coordination
Verification methods for self-organizing systems
Integration testing
Formal proof
Statistical experimental verification
4 Methods of machine learning
Reinforcement learning
Learning classifier systems
Neural Networks
Evolutionary Algorithms
Swarm Intelligence (ACO, PSO)
Explain a k-bit multiplexer
o k = total length, l = length of address bits
o k = l + 2l
o l addresses from right to left [...,2,1,0]
o 2k possible input strings, 2k+l possible rules
o Rule: input:output → reward / payoff
o Input may contain wildcard #
o → Knowledge from trying is called Population
Whats the process of a learning classifier system
1. Detect state of environment
2. Find matching conditions --> match set
3. Select best action (group) --> action set
4. apply output action
5. allocate credit to action set rules
What defines the zeroth-level-classifier system (ZCS)
- Rules with identical action that mach same conditions are competing
- These rules share the received payoff
What defines the eXtended Classifier System (XCS)
Rules are represented by:
- Condition, Action
- Prediction
- Prediction error
- Fitness (inverse prediction error)
- Experience
Evolutionary Algorithms Process
1. Generate initial collection
2. Select parents
3. recombine + mutate
4. Offspring replaces old population
Binary recombination methods
Name 4 selection mechanisms
1. Fitness proportional selection
2. Tournament selection
3. Uniform selection
4. Rank based selection
4 reproduction schemes
generation reproduction [generation replace]
steady-state [single child replace]
(μ,λ)-selection [Sparta]
(μ+λ)-selection [Sparta w. incest]
How can weights for neural networks be obtained
Suvervised training (static problems)
- Hopfield NNs
- Backpropagation NNs
Unsupervised methods (dynamic problems)
- Self-organizing maps (SOM)
- Feature maps
Process of standard ACO
1. For given state i, select action j based on pij formula (above)
2. For selection use roulette wheel
3. Evaporate pheromones
3. Increase pheromone trail on used actions/paths
4. Repeat until stopping criterion is met.
ACO - Design decisions
weights (α , β)
initial pheromone value
evaporation rate p
minimal pheromone value
number of ants per iteration
ants who are allowed to update
stopping criterion
Compare Genetic Algorithms with ACO
- solution-based (solution space)
- local information
- individual solutions are propagated into next generation

- decision based
- global and local information
-information about individual solution not being stored
- constructing new solutions is more time-consuming
What says Robustness , weakly, strongly robust
- Robust: Plateau on parable top : variation in performance function is minimal
- Strongly robust: every possible δ maps TS→ TS and AS→ AS
- Weakly robust: every possible δ maps AS → AS
What is Adaptiveness
- adaptable: every δ maps AS→ SS, external control can lead SS→ AS
- adaptive: adaptable + return to AS without external control
What means Flexiblility
- System is adaptive when objectives changes (and therefore TS and AS change)
What is Reliability
- A system is most reliable where it has the minimal chance fail - leave feasible region
- e.g. in robotics, the best performance could comprise hurting humans.
- hurting humans is not in the feasible region though, so the most reliable solution is not the most performant solution.
What kinds of feedback where discussed
- Robustness needs feedback
- Feedback is information about the result of an action
- Types of Feedback
o positive feedback (leads to cumulative effects)
o negative feedback (leads to equilibrium)
What Elements has Model Predicitive Control
o Model
o Predictive controller
o Process
o Disturbance
Whats the Process of Model Predicitve Control
o 1. Sample output of process
o 2. check if model needs to be updated (wrong prediciton)
o 3. Use model to predict future behaviour (in a prediction horizon)
o 4. Calculate optimal control input that minimalizes the difference between reference output (through objectives) and predicted output
o 5. Apply the input and repeat procedure for the next sample
What are the 8 facets of trustworthiness
What does CIA stand for
--> trusted computing often reduced to these facets
Which approaches of trust management are known?
Credential basted trust
Reputation based trust
Draw the Controller of O/C Architecture
Draw the Observer of O/C Architecture
Draw the basic structure of O/C Architecture
Illustrate the learning classifier system