Quick reference
Glossary: statistics, decisions and learning from A to Z
100 terms from Bayesian statistics, decision theory, simulation, optimisation and reinforcement learning, with concise definitions and links to the full articles.
(
(mixed-)integer programming(mixed-)integer programming, MIP
Optimisation in which some or all variables must be integers, such as item counts or yes/no decisions. These problems are generally NP-hard; branch and bound with LP relaxations is a standard technique.
Articles: Optimization: linear programs and scheduling See also: Branch and bound, NP-hard
A
A/B testA/B test, online controlled experiment
A randomised experiment that assigns users at random to two or more variants to measure the effect of a change on a target metric. Random assignment allows differences to be attributed causally to the variant.
Articles: Bayesian A/B testingMulti-armed bandits & Thompson sampling See also: Conversion rate, Uplift, Multi-armed bandit
Attention
A weighted mixture of value vectors, with weights calculated from query and key vectors. In a transformer, this operation connects information across positions.
Articles: Transformer attention
B
Base rate
The underlying frequency of a characteristic in a population, such as disease prevalence. Ignoring it leads to the base-rate fallacy, confusing with .
Articles: Bayesian statistics explained See also: Prior distribution
Bellman equation
A recursive relationship for value functions: a state’s value is the expected immediate reward plus the discounted value of the next state. The optimality version maximises over actions and characterises the optimal policy.
Articles: Reinforcement Learning explained See also: Value function, Value iteration
Beta distribution
A distribution on with density proportional to and mean . The standard prior for success rates because it is conjugate to the binomial likelihood.
Articles: Bayesian statistics explainedBayesian A/B testingMulti-armed bandits & Thompson sampling See also: Conjugate prior
Bootstrap
Repeated sampling with replacement from an observed sample. Recalculated statistics approximate sampling uncertainty; this does not repair biased data collection.
Branch and bound
A method for integer problems that splits the problem into subproblems (branching). Bounds from relaxations rule out subtrees that cannot improve on the best known solution (bounding).
Articles: Optimization: linear programs and scheduling See also: (mixed-)integer programming
Brier score
For binary outcomes, the mean squared difference between predicted probability and observed outcome. Lower is better. The score evaluates more than calibration alone.
Articles: Calibration: can predicted probabilities be trusted?
Business intelligence
Preparing and presenting business data in reports, metrics and dashboards. It describes what happened and the current situation but does not itself derive an action.
Articles: What is Decision Intelligence? See also: Descriptive analytics, Decision intelligence
C
Calibration
A property of probabilistic forecasts and estimates in which stated probabilities agree with observed frequencies: over time, about 80% of all 80% intervals should contain the actual value. A substantially lower proportion indicates overconfident forecasts.
Articles: What is Decision Intelligence? See also: Prediction interval
Causal effect
Difference between outcomes under two interventions, such as training versus no training. A group difference in observational data is not a causal effect without additional assumptions.
Articles: Causal inference: separating effects from associations
Central limit theorem
A probability theorem: a suitably standardised sum of many independent random variables with finite variance is approximately normally distributed under the usual conditions. A mean’s error is therefore approximately normal with standard deviation .
Articles: Monte Carlo simulation explained See also: Law of large numbers, Standard error
Certainty equivalent
The certain amount a decision-maker regards as equivalent to a risky alternative: , where CE denotes the certainty equivalent. The gap between expected value and certainty equivalent is the risk premium.
Articles: Decision theory and decision trees See also: Risk aversion
Confounding
Distortion of an observed association by common causes of an exposure and an outcome. Appropriate adjustment requires measured relevant causes and justified modelling assumptions.
Articles: Causal inference: separating effects from associations
Conjugate prior
A prior whose posterior belongs to the same distribution family. A beta prior and binomial likelihood, for example, produce posterior . Conjugate pairs allow updates without numerical integration.
Articles: Bayesian statistics explained See also: Beta distribution
Constraint
A condition a feasible solution must meet, such as a capacity limit or minimum quantity. Hard constraints must never be violated; soft constraints are traded off through penalty terms in the objective.
Articles: Optimization: linear programs and scheduling See also: Objective function, Feasible region
Contextual bandit
An extension of the multi-armed bandit that observes context before each choice, such as device, time of day or interests. It learns which option works best in each context; LinUCB is a well-known method.
Articles: Multi-armed bandits & Thompson sampling See also: Multi-armed bandit
Conversion rate
The proportion of visitors or users who perform a desired action, such as a purchase or registration: conversions divided by visitors. In a Bayesian A/B test, each variant’s rate is represented by a beta distribution.
Articles: Bayesian A/B testing See also: A/B test, Beta distribution
Credible interval
An interval containing the parameter with a specified posterior probability, such as 95%. Common choices are the central interval between the 2.5th and 97.5th percentiles and the highest density interval (HDI). Unlike a confidence interval, it makes a direct probability statement, given the model and prior.
Articles: Bayesian statistics explainedBayesian A/B testing See also: Posterior distribution, P-value
D
Decision intelligence
An approach that makes the decision itself the object of design: signals, context, action options, evaluation, approval and feedback from observed outcomes form a verifiable process. It combines statistics, simulation, optimisation and learning from feedback.
Articles: What is Decision Intelligence? See also: Prescriptive analytics, Decision quality, Decision theory
Decision quality
The quality of a decision judged using knowledge available at the time: were relevant options, uncertainties, evaluations and constraints properly considered? It differs from the outcome, which also depends on chance; confusing the two is the fallacy known as resulting.
Articles: What is Decision Intelligence?Decision theory and decision trees See also: Decision rule, Decision intelligence
Decision rule
A rule specified in advance that maps data or metrics to an action, such as choosing a variant once its expected loss falls below a tolerance. Fixed rules make decisions verifiable and guard against justification after the event.
Articles: What is Decision Intelligence?Bayesian A/B testing See also: Expected loss, Decision quality
Decision theory
The study of choosing actions under uncertainty. Actions, possible states of the world, probabilities and evaluations through utility or loss are combined into a justified choice, classically by maximising expected utility.
Articles: Decision theory and decision treesWhat is Decision Intelligence?
Decision tree
A graphical model of a multistage decision, with decision nodes (squares), chance nodes (circles) and terminal outcomes. Solve it backwards: use expected value or expected utility at chance nodes and choose the best option at decision nodes.
Articles: Decision theory and decision trees See also: Value of information (EVPI, Expected value
Deep reinforcement learning
Reinforcement learning with a policy or value function represented by deep neural networks. It enables learning from high-dimensional inputs such as images but usually requires many interactions.
Articles: Reinforcement Learning explained See also: Reinforcement learning, Q-learning
Descriptive analytics
Describing historical and current data through totals, rates, trends and distributions. It answers “What happened?” and provides the foundation for further analysis.
Articles: What is Decision Intelligence? See also: Business intelligence, Predictive analytics
Discount factor
A factor between 0 and 1 that discounts future rewards per time step. Small values make an agent short-sighted; values near 1 give it a longer horizon.
Articles: Reinforcement Learning explained See also: Reward
Discrete-event simulation
Simulation in which state changes at individual event times, such as a job arrival or service completion. There is no need to calculate artificial time steps between events.
Duality
Every linear programme has a dual programme whose variables can be interpreted as resource prices. When finite optimal solutions exist, the primal and dual optimal objective values agree: strong duality.
Articles: Optimization: linear programs and scheduling See also: Shadow price
E
Embedding
A numerical vector representation of an object such as a token or text passage. The similarities reflected by distances depend on the model, its training data and the distance measure.
Articles: RAG and sources
Epsilon-greedy
A simple bandit strategy: choose the option with the highest observed rate with probability , and a random option with probability . With fixed , exploration continues indefinitely without targeting promising options.
Articles: Multi-armed bandits & Thompson sampling See also: Exploration vs. exploitation, Thompson sampling
Expected improvement
A Bayesian optimisation acquisition function: the expected positive improvement over the best value so far, calculated under the predictive model. It balances promising values with uncertainty.
Articles: Bayesian optimisation: which experiment comes next?
Expected loss
The average loss accepted when choosing one option even though another might be better. In an A/B test, choosing B has expected loss . The decision is acceptable when this falls below a threshold chosen in advance.
Articles: Bayesian A/B testing See also: A/B test, Posterior distribution
Expected utility
The probability-weighted mean utility of outcomes, . The von Neumann–Morgenstern theorem represents preferences satisfying specified consistency axioms by maximising expected utility.
Articles: Decision theory and decision trees See also: Utility function, Expected value
Expected value
The probability-weighted mean of a random quantity, . For repeated decisions it corresponds to the long-run average, but says nothing about the spread of outcomes.
Articles: Decision theory and decision trees See also: Expected utility
Exploration vs. exploitation
The fundamental tension in learning systems: exploitation uses the best option according to current knowledge; exploration tries less familiar options to improve that knowledge. Too little exploration misses good options, while too much sacrifices returns.
Articles: Multi-armed bandits & Thompson samplingReinforcement Learning explained See also: Multi-armed bandit, Thompson sampling
F
Feasible region
The set of all solutions satisfying every constraint. In linear programming it is a convex polyhedron; if empty, the problem is infeasible.
Articles: Optimization: linear programs and scheduling See also: Constraint, Linear programming (LP)
G
Gaussian process
A distribution over functions whose values at any finite set of inputs have a joint normal distribution. A covariance function describes how values at different inputs relate.
Articles: Bayesian optimisation: which experiment comes next?
Governance
Rules, roles and controls specifying who owns a decision system, who approves recommendations, how risks are measured and how decisions can be traced. Frameworks such as the NIST AI Risk Management Framework structure these tasks.
Articles: What is Decision Intelligence? See also: Human in the loop
H
Heuristic
A method finding good, though not provably optimal, solutions within reasonable time, such as priority rules, local search or simulated annealing. Bounds from exact methods show the maximum possible gap to optimality.
Articles: Optimization: linear programs and scheduling See also: Scheduling
Holdout
Data withheld when fitting a model. Selecting model variants using these results makes them validation data; an independent final evaluation then needs further test data.
Articles: Regression and overfitting: training is not testing
Human in the loop
A design principle in which a person reviews, approves or corrects a system’s recommendation before it takes effect. It assigns responsibility clearly and requires the recommendation’s assumptions to be visible.
Articles: What is Decision Intelligence? See also: Governance
K
K-means
A method that partitions data into k groups and reduces the sum of squared distances to their centres. Scaling and initial centres affect the result; a global optimum is not guaranteed.
L
Law of large numbers
A probability theorem: the average of independent, identically distributed random variables approaches their expected value as sample size grows. It explains why Monte Carlo estimates improve with more runs.
Articles: Monte Carlo simulation explained See also: Central limit theorem
Likelihood
The probability or density of observed data as a function of the parameter, . With the data fixed, it shows how well each parameter value explains the observations; its integral over generally does not equal 1.
Articles: Bayesian statistics explained See also: Prior distribution, Posterior distribution
Linear programming (LP)
Optimising a linear objective subject to linear equalities and inequalities, such as subject to , . For a feasible, bounded problem in this standard form, an optimum lies at a vertex of the feasible region.
Articles: Optimization: linear programs and scheduling See also: Objective function, Constraint, Simplex method
M
Marginal likelihoodmarginal likelihood, evidence
The overall probability of the data under a model, . It is the normalising factor in Bayes’ theorem and has a separate role in model comparison through Bayes factors.
Articles: Bayesian statistics explained See also: Likelihood
Markov chain Monte Carlo (MCMC)
A family of simulation methods generating a Markov chain with the posterior as its limiting distribution. Frequencies of chain values approximate the posterior. Examples include Metropolis–Hastings, Gibbs sampling and Hamiltonian Monte Carlo (NUTS).
Articles: Bayesian statistics explainedMonte Carlo simulation explained See also: Posterior distribution
Markov decision process (MDP)
A mathematical model for sequential decisions: states, actions, transition probabilities , rewards and a discount factor. The Markov property means the current state contains all information relevant to the future.
Articles: Reinforcement Learning explained See also: Reward, Discount factor, Policy
Maximum a posteriori estimate
The posterior mode: the parameter value with the highest posterior density. With a flat prior, it agrees with the maximum likelihood estimate.
Articles: Bayesian statistics explained See also: Posterior distribution
Minimax regret
A decision rule for situations without probabilities: for each state, measure regret as the gap to that state’s best outcome, then choose the action with the smallest worst-case regret.
Articles: Decision theory and decision trees See also: Payoff table
Monte Carlo simulation
A method estimating an uncertain quantity by evaluating a model repeatedly with randomly drawn inputs and analysing the results statistically. Estimation error decreases in proportion to , where is the number of runs.
Articles: Monte Carlo simulation explained See also: Standard error, Variance reduction, Markov chain Monte Carlo (MCMC)
Multi-armed bandit
A decision problem involving repeated choices between options, or arms, with unknown success rates, observing only the chosen option’s outcome. The goal is to maximise total reward, requiring learning and exploitation simultaneously.
Articles: Multi-armed bandits & Thompson sampling See also: Regret, Exploration vs. exploitation, Thompson sampling
N
NP-hard
A class of problems for which no algorithm is known that solves every instance exactly in polynomial time. Large instances are therefore tackled with problem-specific exact methods and bounds, or with heuristics.
Articles: Optimization: linear programs and scheduling See also: Heuristic, (mixed-)integer programming
O
Objective function
The quantity an optimisation model maximises or minimises, such as contribution margin, cost or delay. Which solution is “optimal” depends entirely on how the objective is formulated.
Articles: Optimization: linear programs and scheduling See also: Constraint
Optional stoppingoptional stopping, peeking
Ending an experiment based on results observed so far, for example when a metric crosses a threshold. The Bayesian posterior remains correctly calculated, but repeated peeking changes a decision rule’s error rates across many experiments.
Articles: Bayesian A/B testing See also: A/B test, P-value
Overfitting
A model fits quirks and noise in its training data so closely that it performs worse on new data. Low training error alone therefore does not establish predictive quality.
Articles: Regression and overfitting: training is not testing
P
P-value
The probability, under the null hypothesis, of observing a result at least as extreme as the actual one. It is not the probability that the null hypothesis is true and does not measure effect size.
Articles: Bayesian statistics explainedBayesian A/B testing See also: Credible interval
Parameter
An unknown quantity in a statistical model, such as a success rate or mean . Bayesian statistics represents a parameter through a probability distribution.
Articles: Bayesian statistics explained
Payoff table
A table giving the outcome for each combination of action (row) and state of the world (column). The starting point for decision rules such as maximin, minimax regret and expected value.
Articles: Decision theory and decision trees See also: Decision theory, Minimax regret
Policy
An agent’s strategy: a rule mapping each state to an action or a probability distribution over actions.
Articles: Reinforcement Learning explained See also: Value function
Posterior distribution
A parameter’s distribution after accounting for data: proportional to likelihood times prior, . It provides point estimates, credible intervals and probabilities of statements.
Articles: Bayesian statistics explainedBayesian A/B testing See also: Prior distribution, Likelihood, Credible interval
Posterior predictive distribution
The distribution of future observations, averaged over all plausible parameter values: . Used for predictions and posterior predictive model checks.
Articles: Bayesian statistics explained
Prediction interval
An interval into which a future observation falls with a specified probability, such as 80%. Unlike a confidence or credible interval for a parameter, it concerns a future individual value and is therefore wider.
Articles: What is Decision Intelligence? See also: Calibration, Credible interval
Predictive analytics
Using statistical models or machine learning to estimate what is likely to happen under stated assumptions. For decisions, a forecast distribution with a prediction interval is more valuable than a single number.
Articles: What is Decision Intelligence? See also: Prediction interval, Prescriptive analytics
Prescriptive analytics
Deriving recommended actions from forecasts, objectives and constraints, usually through optimisation, simulation or decision rules. It requires explicitly stated objectives and boundaries.
Articles: What is Decision Intelligence?Optimization: linear programs and scheduling See also: Decision intelligence, Decision rule
Prior distribution
A probability distribution describing which values of an unknown parameter are plausible before looking at the data. It may be flat, weakly informative or informative and should be justified and documented.
Articles: Bayesian statistics explained See also: Posterior distribution, Conjugate prior
Pseudo-random numbers
Number sequences generated by a deterministic algorithm that behave statistically like randomness. The same starting value, or seed, produces the same sequence, making simulations reproducible and verifiable.
Articles: Monte Carlo simulation explained
Q
Q-learning
A model-free reinforcement learning method that learns action values from experience: . It learns the optimal policy’s values even while the agent explores: off-policy learning.
Articles: Reinforcement Learning explained See also: Temporal-difference learning, Bellman equation
Quantilequantile, percentile
A value below which a specified proportion of a distribution lies. The 90th percentile, or P90, is not exceeded in 90% of cases; the median is the 50th percentile.
Articles: Monte Carlo simulation explained See also: Triangular distribution
Queue utilisation
Offered workload relative to service capacity. In the simple model, ρ = arrival rate × mean service time / number of servers; at ρ ≥ 1 its usual assumptions do not allow stable stationary operation.
R
Region of practical equivalence
Region of practical equivalence: a range of effect sizes defined in advance as practically equivalent to zero. An effect is considered relevant if the highest posterior density interval lies entirely outside it, and negligible if entirely inside.
Articles: Bayesian A/B testing See also: Credible interval
Regret
The difference between the expected reward obtainable with the best option and the expected reward actually obtained: . Regret measures learning’s cost; good bandit algorithms make it grow only logarithmically.
Articles: Multi-armed bandits & Thompson sampling See also: Multi-armed bandit
Reinforcement learning
A branch of machine learning in which an agent learns through interaction with an environment to choose actions that maximise the long-term, usually discounted sum of rewards.
Articles: Reinforcement Learning explained See also: Markov decision process (MDP), Q-learning
Retrieval-augmented generation
Combining source retrieval with a generative language model by supplying relevant document passages as additional context. Relevant retrieval, current sources and a supported answer each require evaluation.
Articles: RAG and sources
Reward
A numerical value the environment gives an agent after an action in reinforcement learning. It is the learning signal; its design determines which behaviour is actually learned.
Articles: Reinforcement Learning explained See also: Discount factor
Reward hacking
A learning or optimising system achieves high scores for a specified reward while its behaviour misses the intended goal. An incomplete or exploitable objective can cause this.
Articles: Reward design
Risk aversion
A decision-maker’s preference for a certain amount over a lottery with the same expected value. It corresponds to a concave utility function; the certainty equivalent is then below the expected value.
Articles: Decision theory and decision trees See also: Certainty equivalent, Utility function
S
Sample ratio mismatch (SRM)
A substantial discrepancy between actual and planned user allocation, such as 50.8:49.2 instead of 50:50 in a large sample. Usually indicates a technical problem in assignment or measurement, making test results unreliable.
Articles: Bayesian A/B testing See also: A/B test
Scenario
An internally consistent assumption about the future, such as demand, prices or failures, under which action options are evaluated. Monte Carlo simulation generates many scenarios from probability distributions to show the range of possible outcomes.
Articles: What is Decision Intelligence?Monte Carlo simulation explained See also: Prediction interval
Scheduling
Planning which tasks run on which resources and when, with objectives such as short completion times or low delay. Even basic variants such as job-shop scheduling are NP-hard.
Articles: Optimization: linear programs and scheduling See also: NP-hard, Heuristic
Sensitivity analysis
Investigating how strongly a model’s result reacts to changes in individual inputs or assumptions. Variance-based methods measure the share of output variation attributable to each input.
Articles: Monte Carlo simulation explained See also: Monte Carlo simulation
Shadow priceshadow price, dual value
The improvement in the optimal objective value per additional unit of a scarce resource, given by the corresponding dual variable. Valid only within a range where the same constraints remain binding; unused resources have a zero shadow price.
Articles: Optimization: linear programs and scheduling See also: Duality
Shrinkage
Pulling an estimate towards a prior assumption or an overall mean. In the beta-binomial model, the posterior mean is a weighted average of the prior mean and observed rate, protecting small samples against extreme estimates.
Articles: Bayesian statistics explained
Simplex method
George B. Dantzig’s method for linear programmes, moving between adjacent vertices of the feasible region without worsening the objective value. Exponential in the worst case but usually very fast in practice.
Articles: Optimization: linear programs and scheduling See also: Linear programming (LP)
Softmax
A function that exponentiates and normalises real values into positive weights summing to one. Large weights alone do not establish calibrated probabilities.
Articles: Transformer attention
Standard error
The standard deviation of an estimator across hypothetical repetitions of a sample or simulation. For a mean of independent values it is ; under a normal approximation, roughly 95% of estimates lie within two standard errors of the true value.
Articles: Monte Carlo simulation explained See also: Central limit theorem
Standardisation
Here: transforming a feature by subtracting its mean and dividing by its standard deviation. This reduces the influence of large units on distance calculations; it does not determine substantive relevance.
T
Temporal-difference learning
A family of methods shifting value estimates after each step towards a target formed from the observed reward and estimated next-state value. The difference between the target and current estimate is the TD error.
Articles: Reinforcement Learning explained See also: Q-learning
Thompson samplingThompson sampling, probability matching
A bandit strategy that draws one random value from each option’s posterior and chooses the option with the largest draw. This selects each option with precisely the probability that it is currently the best.
Articles: Multi-armed bandits & Thompson sampling See also: Multi-armed bandit, Posterior distribution, Upper confidence bound
Token
A unit into which a language model divides text, such as a word fragment or punctuation mark. The segmentation depends on the tokenizer; a token is not necessarily a whole word.
Articles: Transformer attention
Triangular distribution
A continuous distribution specified by minimum , most likely value and maximum , with a triangular density and mean . Popular for three-point estimates in project and risk planning.
Articles: Monte Carlo simulation explained See also: Quantile
U
Upliftuplift, relative lift
The relative improvement of a variant over the control, usually . Always report an interval alongside the point estimate, because uncertainty about the size is often substantial, especially for small effects.
Articles: Bayesian A/B testing See also: A/B test, Credible interval
Upper confidence bound
A family of bandit strategies based on optimism under uncertainty: each option is evaluated using an upper confidence bound on its rate. UCB1 selects the arm with the largest , giving rarely played arms a bonus.
Articles: Multi-armed bandits & Thompson sampling See also: Thompson sampling, Epsilon-greedy
Utility function
A function assigning a utility value to outcomes, such as monetary amounts. Its curvature represents risk attitude: concave for risk aversion, linear for risk neutrality and convex for risk seeking.
Articles: Decision theory and decision trees See also: Risk aversion, Expected utility
V
Value function
The expected future return when starting in a state (state value ), or first choosing a particular action there (action value ), then following policy .
Articles: Reinforcement Learning explained See also: Bellman equation, Policy
Value iteration
A dynamic programming method repeatedly applying the Bellman optimality equation as an update until values stop changing. It requires known transitions and rewards.
Articles: Reinforcement Learning explained See also: Bellman equation, Q-learning
Value of information (EVPIvalue of information (EVPI, EVSI)
How much the best achievable expected value improves when additional information is received before deciding. The expected value of perfect information (EVPI) bounds the value of any study; the value of a particular sample or test is EVSI.
Articles: Decision theory and decision trees See also: Decision tree
Variance reduction
Techniques reducing the variability of a Monte Carlo estimate for a fixed number of runs, such as antithetic variables, common random numbers, control variates and importance sampling.
Articles: Monte Carlo simulation explained See also: Monte Carlo simulation, Standard error