Exam Flashcards
Genetic Algorithmus
Give three selection Methodes to constantly improve the population
Elitism:
Copies the best chromosomes to the new population before loosing them through crossover or mutation
Roulette Wheel Selection:
Size of wheel section according to fitness value
Rank selection:
Like Roulette but wheel section size depends on fitness value ranking
Genetic Algorithms
Which three forns of encoding exist and briefly explain how they work
Binary Encoding:
Every chromosome is a string of bits(0,1)
Permutation Encoding:
Every chromosome is a string of numbers, which represents the order of a sequence
Value Encoding:
Every chromosome is a string of values. Values can be anything
Genetic Algorithms
Give a pseudo-code for a GA-program
Start Encoding Generate initial population Compute fitness Until population has concerged: selection crossover mutation compute fitness stop
Genetic Algorithms
What are the basic terms in GA
population
chromosome
gene
Genome: Complete set of genetic material
Genotype: Particular set of genes in genome
Fuzzy Logic
Name the 4 Fuzzy Logic controller types
Mandani
Tsukamoto
TSK
Laroen
Fuzzy Logic
Give the formal representation of a fuzzy set.
Give the operations on membership functions with their respective formulas and draw name an exemplary mf
Support: Non zero membership degree
Boundary: membership degrees 0 < my_A(x) < 1
Core: membership degree 1
Singletion: supp(A) = core(a) = {x_0}
Alpha cut my_A(x) >= alpha
Strong alpha cut set_ my_A(x) > alphaFuzzy Logic
How are fuzzy and crisp sets defind and how do boolean logic and fuzzy logic stand to each other?
Fuzzy set:
Sets whose elements have degrees of membership
Crisp set:
degree of membership is either 1 or 0
Boolean logic is a special case of fuzzy logic
Q-Learning & SARSA
Explaun the difference between Q-Learning and SARSA
SARSA
updates with respekt to the result of the action taken
leans values of policy carried out
Q
assumes optimal policy is beeing followes
leans optimal policy independent of agent’s actions
Q-Learning
Gice the update value for Q-Learning and explain basic features of Q-Learning.
mapping (state, action) to reward all features values on optimal path
model free learning
Q(s,a) converges to the expected value of a state when following an optimal policy
off policy learning
agent always takes optimal action
SARSA
Geive the update rule for SARSA and explain basic features
on policy learning
agant has free choice -> non-optimal action
Reinforcement Learning
Define the value function v(s) and give the formulas to the two models for value function
- The agent’s current belive how good a state is, given a sll expected future rewad. The value of a state is defined as the sum off all expected future rewards, assuming we are in state s and follow pi(s)
Finite Horizon Model
Infinite Horizon Model
Reinforcement Learning
Which tyes of rewards are there
- Pure delayed award
- Minimum time to goal award
- Multi player games
Auto-asscociative nn (autoencoder)
Outline an auto-assoc. nn and explain how it works
- performs an identity mapping (i=o)
- in the middle of the network is a bottleneck-layer that enforces a reduction of dimension of the data
- if lin function in the center it will find the first principle component (direction in which data has lowest variance)
Hopefield Network auto associative memories
How can such a network remember patterns?
How can more than one pattern be learned?
- nn stores and retrives associations as synaptic connection. Hebbina learning presynaptic and postsynaptic neurons are envolved. ass. or content addressable way (STDP)
- Out. of each neuron is fed back via unit-time delay, to all other neurons in the network. NN can recognize patterns through only partial exposure of the pattern.
asynchronous and synchronous update possible - add up all weight patterns:
w = s_taret * s_input
s_past m= sigma(w*s_present)
patterns/neurons > 0.138 otherweise small errors pile up
Kohonen’s self-organized-maps (SOM)
Compare a SOM to VQ and name the biggest difference between them. Give and explain two typical practical problems.
SOM
- better at overcoming the under- or overfitting and local minima problem
- produces map with some ordering among the node vectors and this gives the map the ability to tolerate noise in the input or retival patterns
- use of neighbourhood function
- (bad) mapping into 2D-weights -> knots can appear
- (bad) mapping intp a too-low dimension
