Decision model
How it works
1) Define the decision space: the set of admissible actions/alternatives. 2) Represent state and uncertainty: input features and probabilities of states of nature (often from a predictive model). 3) Objective function: utility, cost, or risk assigned to (action, outcome) pairs. 4) Decision criterion: e.g. maximize expected utility, minimize expected cost, or min-max strategies when no distribution is available. 5) Rule/policy: a mapping from state to action, derived analytically (e.g. a Bayes threshold), by dynamic programming (MDP), or learned (an RL policy). 6) Execution and update: the decision is taken, the outcome observed, and the model may be refined (e.g. online learning).
Problem solved
It solves the problem of moving from knowledge about an uncertain future to a concrete action: a probability alone does not say what to choose when error costs are asymmetric or constraints exist. A decision model explicitly encodes the objective, costs, and rules, making decisions consistent, repeatable, and justifiable.
Components
The set of admissible actions or alternatives from which the model chooses.
Input features and the probabilities of states of nature, usually supplied by a predictive model.
A valuation of action-outcome pairs encoding preferences, risk, and the asymmetry of error costs.
A mapping from state to action, derived analytically, by dynamic programming, or learned.
Evolution
von Neumann and Morgenstern formalize rational choice via expected-utility maximization — the foundation of decision models.
Bellman and dynamic programming provide a method for computing optimal policies in sequential decisions.
Deep reinforcement learning (e.g. DQN) makes it possible to learn decision models directly from raw observations.
Hyperparameters (configurable axes)
The optimization rule: expected utility, risk minimization, min-max, satisficing.
Whether decisions are discrete (a finite set) or continuous — determines the optimization method.
Hardware requirements
A decision model is a logical/mathematical layer; its cost depends on the underlying predictive model or optimization method, not on a specific accelerator.