Karteikarten
08 Bayesian Networks
Karten in diesem Stapel (7)
- 1Which motivation stands behind Bayesian networks?Acting rationally given uncertain information
Used to represent dependencies among variables to reduce number of information required to construct joint probability - 2How are Bayesian networks structured?Acyclic graph, where:
-each node corresponds to a random variable
-arrow between nodes starts at parents
-each node has a conditional probability distribution - 3What are inference tasks here?-Simple queries: compute probabilities given some evidence
-Conjunctive queries
-Optimal decisions (outcome I action, evidence)
-Value of information
-Sensitivity analysis
-Explanation - 4Which kinds of inferences can be applied?-Inference by enumeration: sum out variables from joint probability without actually constructing its explicit representation, recursive depth-first - inefficient
-Inference by variable elimination: storing to avoid re-computation, pointwise product & summing out variables - 5What is an alternative to the Monte Carlo Simulation and which are its two types?Direct sampling:
-Rejection sampling
-Likelihood weighting - 6What are characteristics of rejection sampling?-produce samples from a hard-to-sample distr. given an easy-to-sample distr.
-determining conditional probabilities
-consistent posterior estimates
-very expensive if P(e) is small (P(e) drops off exponentially with no. of evidence var.) - 7What are characteristics of likelihood weighting?-avoids inefficiency by only generating events incl. evidence e (=importance sampling) and as not all events are equal, weight each of them by likelihood of occurence
-sample only non-evidence variables
-performanc still degrades with many ev. var. as a few samples have nearly all the weighting
Informationen
- Beschreibung
- Artificial Intelligence Karteikarten am 08 Bayesian Networks, erstellt von Timo Bierwirth am 22/02/2018.
- Erstellt von
- Timo Bierwirth
- Aufrufe
- 0
- Bewertung