Uncertainty is incomplete knowledge about a state, quantity, or outcome. A learning or decision problem begins with something unresolved, uses available evidence to produce an estimate, probability, or decision, and remains limited by assumptions, data, and computation. section 10.0 section 13.4.1
Different methods address uncertainty in different ways. Bayesian inference revises beliefs as evidence arrives; maximum likelihood estimates parameters from observed data; and hypothesis testing uses decision rules that account for the possibility of error. These methods need not eliminate uncertainty to support a useful conclusion. section 5.0 section 4.5
Information theory makes uncertainty measurable. Shannon entropy describes the average uncertainty associated with a source before an outcome is observed, expressed in bits. An individual outcome conveys information according to how surprising it is: a predictable outcome generally conveys less than an unlikely one. Entropy is an average over possible outcomes, not the information conveyed by any one outcome. section 5.1.1 section 5.6
Fair and biased coins
Before a fair coin lands, heads and tails are equally likely; observing either outcome resolves which occurred. With a coin that lands heads 90% of the time, heads is less surprising and conveys less information than the unlikely tails outcome. section 5.1.1
Uncertainty can concern an outcome even when its governing model is known, or it can concern the model itself. In a Markov decision process with known transition probabilities and rewards, dynamic programming evaluates uncertain outcomes using that model. When those probabilities or rewards are unknown, reinforcement learning can instead use experience to improve decisions. section 6.5.3
A positive diagnostic test
A positive test can be informative without being conclusive. With a disease base rate below 1%, the result raises the probability of disease to approximately 9%; imperfect testing and limited further evidence leave substantial uncertainty. section 10.2.1
Reducing uncertainty is not the same as guaranteeing a correct decision. In hypothesis testing, acceptable Type I and Type II error risks are specified before data analysis. In prediction, a logistic model outputs a probability, which may then be converted to a class label using a threshold; the label does not erase the uncertainty expressed by the probability. section 4.5 section 3.3.2
















