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.

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.

Attention

A weighted mixture of value vectors, with weights calculated from query and key vectors. In a transformer, this operation connects information across positions.

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 P(A∣B)P(A \mid B) with P(B∣A)P(B \mid A).

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.

Beta distribution

A distribution on [0,1][0, 1] with density proportional to θα−1(1−θ)β−1\theta^{\alpha-1}(1-\theta)^{\beta-1} and mean α/(α+β)\alpha/(\alpha+\beta). The standard prior for success rates because it is conjugate to the binomial likelihood.

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).

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.

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.

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.

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.

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 σ/N\sigma/\sqrt{N}.

Certainty equivalent

The certain amount a decision-maker regards as equivalent to a risky alternative: u(CE)=E[u(X)]u(\text{CE}) = \mathbb{E}[u(X)], where CE denotes the certainty equivalent. The gap between expected value and certainty equivalent is the risk premium.

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.

Conjugate prior

A prior whose posterior belongs to the same distribution family. A beta prior and binomial likelihood, for example, produce posterior Beta⁡(α+k,β+n−k)\operatorname{Beta}(\alpha+k, \beta+n-k). Conjugate pairs allow updates without numerical integration.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

Descriptive analytics

Describing historical and current data through totals, rates, trends and distributions. It answers “What happened?” and provides the foundation for further analysis.

Discount factor

A factor γ\gamma 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.

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.

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.

Epsilon-greedy

A simple bandit strategy: choose the option with the highest observed rate with probability 1−ε1-\varepsilon, and a random option with probability ε\varepsilon. With fixed ε\varepsilon, exploration continues indefinitely without targeting promising options.

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.

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 E[max⁡(pA−pB,0)]\mathbb{E}[\max(p_A - p_B, 0)]. The decision is acceptable when this falls below a threshold chosen in advance.

Expected utility

The probability-weighted mean utility of outcomes, ∑ipi u(xi)\sum_i p_i\,u(x_i). The von Neumann–Morgenstern theorem represents preferences satisfying specified consistency axioms by maximising expected utility.

Expected value

The probability-weighted mean of a random quantity, E[X]=∑ipi xi\mathbb{E}[X] = \sum_i p_i\,x_i. For repeated decisions it corresponds to the long-run average, but says nothing about the spread of outcomes.

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.

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.

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.

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.

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.

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.

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.

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.

Likelihood

The probability or density of observed data as a function of the parameter, p(y∣θ)p(y \mid \theta). With the data fixed, it shows how well each parameter value explains the observations; its integral over θ\theta generally does not equal 1.

Linear programming (LP)

Optimising a linear objective subject to linear equalities and inequalities, such as max⁡ c⊤x\max\, c^\top x subject to Ax≤bAx \le b, x≥0x \ge 0. For a feasible, bounded problem in this standard form, an optimum lies at a vertex of the feasible region.

M

Marginal likelihoodmarginal likelihood, evidence

The overall probability of the data under a model, p(y)=∫p(y∣θ) p(θ) dθp(y) = \int p(y \mid \theta)\,p(\theta)\,\mathrm{d}\theta. It is the normalising factor in Bayes’ theorem and has a separate role in model comparison through Bayes factors.

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).

Markov decision process (MDP)

A mathematical model for sequential decisions: states, actions, transition probabilities P(s′∣s,a)P(s' \mid s, a), rewards and a discount factor. The Markov property means the current state contains all information relevant to the future.

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.

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.

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 1/N1/\sqrt{N}, where NN is the number of runs.

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.

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.

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.

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.

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.

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.

Parameter

An unknown quantity in a statistical model, such as a success rate θ\theta or mean μ\mu. Bayesian statistics represents a parameter through a probability distribution.

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.

Policy

An agent’s strategy: a rule π(a∣s)\pi(a \mid s) mapping each state to an action or a probability distribution over actions.

Posterior distribution

A parameter’s distribution after accounting for data: proportional to likelihood times prior, p(θ∣y)∝p(y∣θ) p(θ)p(\theta \mid y) \propto p(y \mid \theta)\,p(\theta). It provides point estimates, credible intervals and probabilities of statements.

Posterior predictive distribution

The distribution of future observations, averaged over all plausible parameter values: p(y~∣y)=∫p(y~∣θ) p(θ∣y) dθp(\tilde y \mid y) = \int p(\tilde y \mid \theta)\,p(\theta \mid y)\,\mathrm{d}\theta. Used for predictions and posterior predictive model checks.

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.

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.

Prescriptive analytics

Deriving recommended actions from forecasts, objectives and constraints, usually through optimisation, simulation or decision rules. It requires explicitly stated objectives and boundaries.

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.

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.

Q

Q-learning

A model-free reinforcement learning method that learns action values from experience: Q(s,a)←Q(s,a)+α [r+γmax⁡a′Q(s′,a′)−Q(s,a)]Q(s,a) \leftarrow Q(s,a) + \alpha\,[r + \gamma \max_{a'} Q(s',a') - Q(s,a)]. It learns the optimal policy’s values even while the agent explores: off-policy learning.

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.

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.

Regret

The difference between the expected reward obtainable with the best option and the expected reward actually obtained: RT=Tμ∗−E[∑trt]R_T = T\mu^* - \mathbb{E}[\sum_t r_t]. Regret measures learning’s cost; good bandit algorithms make it grow only logarithmically.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

Softmax

A function that exponentiates and normalises real values into positive weights summing to one. Large weights alone do not establish calibrated probabilities.

Standard error

The standard deviation of an estimator across hypothetical repetitions of a sample or simulation. For a mean of NN independent values it is σ/N\sigma/\sqrt{N}; under a normal approximation, roughly 95% of estimates lie within two standard errors of the true value.

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.

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.

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.

Triangular distribution

A continuous distribution specified by minimum aa, most likely value bb and maximum cc, with a triangular density and mean (a+b+c)/3(a+b+c)/3. Popular for three-point estimates in project and risk planning.

U

Upliftuplift, relative lift

The relative improvement of a variant over the control, usually pB/pA−1p_B / p_A - 1. Always report an interval alongside the point estimate, because uncertainty about the size is often substantial, especially for small effects.

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 μ^i+2ln⁡t/ni\hat\mu_i + \sqrt{2\ln t / n_i}, giving rarely played arms a bonus.

Utility function

A function uu 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.

V

Value function

The expected future return when starting in a state (state value Vπ(s)V^\pi(s)), or first choosing a particular action there (action value Qπ(s,a)Q^\pi(s, a)), then following policy π\pi.

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.

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.

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.

From NyxAI

smartlytics is the learning hub of NyxAI GmbH.

Explore the methods we work with, explained openly, with sources and calculations you can reproduce.

Decision Core

NyxAI is developing a Decision Core for Decision Intelligence that combines optimisation, simulation, statistics and reinforcement learning: compare options, make uncertainty visible and learn from outcomes, with people in control.

AI workflows at NyxAI

DeepSquare

Explainable chess AI with human playing styles: NyxAI’s testing ground for understandable decisions under time pressure.