Lecture 10: Probability and Statistics (part 2)This WeekSamplingSlide 4Sampling: Simulation OutputEstimating Distribution from SamplesSampling DistributionsPoint EstimationConfidence IntervalsConfidence Intervals in SimulationHypothesis TestsErrors in Hypothesis TestingHypothesis Testing in SimulationCOMP155Computer SimulationOctober 1, 2008This WeekReview of probability and statistics needed to understand simulationfollow Appendix C in Arena textOutlineMonday (C.1, C.2):○Probability – basic ideas, terminology○Random variables, joint distributionsToday (C.3-C.5):○Sampling○Statistical inference – point estimation, confidence intervals, hypothesis testingSamplingStatistical analysis: purpose is to estimate or infer something about a large population population is a set of data pointspopulation is too large to look at completely,so we only look at a sample from the populationif the sample is randomly selected from the population, the distribution of the sample should be the same as the distribution of the populationin practice: determine a PMF or PDF for a sampleand assume that distribution holds for the entire populationSamplingRandom sample is a set of independent and identically distributed (IID) observations X1, X2, …, Xn from the populationInput modeling: ○observations come from the real world○Arena’s input analyzer can be used to determine distribution functionOutput analysis: ○observations are the results of multiple runs/replications of the simulation○Arena’s output analyzer can be used to characterize the output population from the observations.Sampling: Simulation OutputRandom sample is a set of independent and identically distributed (IID) observations X1, X2, …, Xn from the populationInput modeling: ○observations come from the real world○Arena’s input analyzer can be used to determine distribution functionOutput analysis: ○observations are the results of multiple runs/replications of the simulation○Arena’s output analyzer can be used to characterize the output population from the observations.Estimating Distribution from SamplesSamples: X1, X2, …, Xn assuming a normal distribution, compute:sample meansample varianceThese statistics have their own sampling distribution, which is generally normalnXXnii11)(122nXXsniiSampling DistributionsIf If underlying distribution of X is normal, then the distribution of is also normal.Point EstimationPoint estimates are estimates of population distribution parameters (2, …)Properties of point estimates Unbiased: E(estimate) = parameterEfficient: Var(estimate) is lowest among competing point estimatorsConsistent: Var(estimate) decreases (usually to 0) as the sample size increasesConfidence IntervalsA confidence interval quantifies the likely imprecision in a point estimatorAn interval that contains (covers) the unknown population parameter some specified probabilityCalled a 100 (1 – )% confidence interval for the parameterExample: 87 < < 123 with probability 95%The value of is in (87, 123) with 95% confidenceWe’ll leave the computation of confidence intervals to a statistics course … or to Arena’s output analyzer tool.Confidence Intervals in SimulationRun simulation replications, get resultsView each replication of the simulation as a data pointForm a confidence intervalThe confidence interval tells you how close you are to getting the “true” expected output (what you’d get by averaging an infinite number of replications)Hypothesis TestsA hypothesis test is used to test some assertion about the population or its parametersWith sampling, we don’t get true/false result, only get evidence that points one way or anotherNull hypothesis (H0) – what is to be testedAlternate hypothesis (H1 or HA) – denial of H0H0: = 6 vs. H1: 6H0: < 10 vs. H1: 10H0: 1 = 2 vs. H1: 1 2Develop a decision rule to decide on H0 or H1 based on sample dataErrors in Hypothesis Testing H0 is really true H1 is really true Decide H0 (“Accept” H0) No error Probability 1 – is chosen (controlled) Type II error Probability is not controlled – affected by and n Decide H1 (Reject H0) Type I Error Probability No error Probability 1 – = power of the test 1-α is the probability of your confidence intervalHypothesis Testing in SimulationInput sideSpecify input distributions to drive the simulationCollect real-world data on corresponding processes“Fit” a probability distribution to the observed real-world dataTest H0: the data are well represented by the selected distributionOutput sideHave two or more “competing” designs modeledTest H0: all designs perform the same on output, or test H0: one design is better than anotherSelection of a “best” model
View Full Document