Scribd
Upload
681 views

CS 188 Introduction To AI Midterm Study Guide

Variable elimination is an algorithm for inference in Bayesian networks that operates by repeatedly selecting and eliminating hidden variables. It involves joining factors containing the variable, marginalizing over the variable to eliminate it, and normalizing the resulting factor. The complexity depends on the size of the largest factor. Gibbs sampling is a Markov chain Monte Carlo method for estimating conditional probabilities in Bayesian networks. It works by randomly initializing a state consistent with evidence and then repeatedly resampling non-evidence variables based on their conditional distributions given the current values of other variables. Hidden Markov models represent systems as a chain-structured Bayesian network where future states depend only on the present state. Particle filtering uses a set of weighted samples, or "particles", to represent the belief state in hidden

Uploaded by

spunkymonkey1016
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
681 views

CS 188 Introduction To AI Midterm Study Guide

Variable elimination is an algorithm for inference in Bayesian networks that operates by repeatedly selecting and eliminating hidden variables. It involves joining factors containing the variable, marginalizing over the variable to eliminate it, and normalizing the resulting factor. The complexity depends on the size of the largest factor. Gibbs sampling is a Markov chain Monte Carlo method for estimating conditional probabilities in Bayesian networks. It works by randomly initializing a state consistent with evidence and then repeatedly resampling non-evidence variables based on their conditional distributions given the current values of other variables. Hidden Markov models represent systems as a chain-structured Bayesian network where future states depend only on the present state. Particle filtering uses a set of weighted samples, or "particles", to represent the belief state in hidden

Uploaded by

spunkymonkey1016
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Inference by Enumeration

1. 2.

Select entries consistent with evidence Sum out H to get joint of Query and evidence

3.

Normalize

Variable Elimination While there are still hidden variables: 1. Pick hidden variable H 2. Join all factors that have H 3. Marginalize H (eliminate by summing out, normalize) new factor Complexity determined by largest factor. Doesnt always exist an ordering that only results in small factors. NP-hard in BN.
A path is active if each triple is active. Want to check if X and Y are conditionally independent given Z. If there exists an active path between X and Y, then cond. Indep. not guaranteed. If path is inactive, X indep of Y. Prior Sampling: Start at top of Bayes Net and work downward. Sample each variable V based on P(V | parents). Retain all samples. Rejection Sampling We want to estimate P(Query | Evidence) Start at top of Bayes Net and work downward. Sample each variable and once we sample a variable not consistent with evidence, we reject that sample. Likelihood Weighting We want to estimate P(Query | Evidence) Start at top of Bayes Net and work downward. If we reach a non-evidence variable, sample it. if we reach an evidence variable, fix it to be consistent with evidence and weigh your sample by P(evidence variable | parents). If we have multiple evidence variables, the weight of your sample would be product of P(E | parents) for each evidence variable E. P(Query) = Sum(wts of samples with Query) / Sum(wts of samples with evidence (all if Query isnt conditional)) Gibbs Sampling We want to estimate P(Query | Evidence) How to get one gibbs sample? Step 1: Randomly choose an instantiation of all variables consistent with evidence. Step 2: Repeat Choose non-evidence variable H, resample it based on P(H | all other evidence), obtaining an updated sample. important: we sample based on all other evidence, not just parents. ^Ideally you would repeat this infinitely many times, but in practice, you would sample a "large amount". Repeat steps 1 and 2 for more gibbs samples. I.e if you want 5 Gibbs samples, do steps 1 and 2 five times. When to use which? If we dont know nothing about probability we are interested in. Prior Sampling: we can use prior sampling, which will be able to answer a variety of queries, priors, joins, etc. We will be able to estimate P(query | evidence) if we have at least one sample consistent with the evidence. If we know the probability we are interested it (i.e evidence for P(query | evidence)): In general, likelihood weighting is better than rejection sampling since we dont reject any parents and the evidence is fixed. Likelihood has its limits as well and it depends on what probability you are interested in and structure of your Bayes net.

Worst-case O(d ) and same for space complexity to store joint distribution

Probability Rules ( ) Product Rule: ( ) ( | ) Chain Rule: ( ) ( | ) ) ( | ) ( ) Bayes Rule: ( ( | ) ( ) ( | ) ( | ) ( ) ( ) ( ) ( ) ( ) X is conditionally independent from Y given Z ( | ) | iff ( | ) ( | ) or (equiv) ( | ) ( | ) VPI properties: Non-negative, non-additive, an order-independent. No such thing as value of ( ) imperfect. In general if | then ( | ) Bayes Nets Assume conditional independence ( | Joint prob of x1xn = ( )) A joint distr over N Boolean variables is 2N An N-node Bayes net with nodes with up to k parents is O(N*2k+1)

Hidden Markov Models and Particle Filtering * Chain-structured BN. Past and future events independent of the present. * Hidden process: future depends on past via present. Evidence vars not guaranteed to be independent. ( ) * Stationary distribution if ( ) ( | )( | ) ( ) * Base cases: Evidence update: ( | ) ( ) ( ) ( ) ( ) ( | ) ( ) Time step: ( ) ( ) ( | ) * In passage of time, beliefs get pushed through the transitions. Current belief ( ) ( | ). one time step passes ( | ( | ) ( | ) ) ) * observation P(X|prev evidence) ( ( | ) then ( ) ( | ) ( ) and renormalize. Forward algorithm combines steps

Particle Filtering elapse time

Observe

Resample

DBN particle filtering

You might also like