This section explains some basic concepts that can help to analyse a decision-making problem at high level.
Decision Science in Context
Decision
science is the problem of finding an algorithmic policy that makes the best possible decision (optimal action) so as to achieve specific objectives subject to specific rules (constraints). Essentially being a optimization problem, Its systematic development has gone through four major phases (so far): (1) starting in the 1940s with cybernetics as studying the control of dynamic systems (i.e. systems with feedback loops); (2) during 1960s with the systematic systems and control of linear dynamic systems, leading to a rich theory of stable control algorithms; (3) during the 1980s with studying the control (and its complexity) of nonlinear dynamic systems, and with neurofuzzy control an early example of an AI-inspired heuristic control (i.e. decision) solution; (4) after 2010 with reinforcement learning providing an algorithmic policy to learning a system model about the behaviour of its environment whilst concurrently trying to control it . Though developments through all stages are based on optimization principles, provable safety guarantees, and risk assessments are much harder to obtain for stages (3) and later, Yet problems typical for these later stages, such as social, economic, health, defence, and biological problems, have immense interest, and only with making the effort to understand their complexities, potential application and risks shall we be able to master and manipulate them, responsibly.
Distributed Decision-Making
Often decisions result from the collaboration of independent decision-makers (i.e. agents); e.g. think of human organizations. This distributed decision-making concept is quite general in that (1) each decision node can potentially be fulfilled by a human or a machine (depending on the level of routine), and (2) may function as a modularization scheme for decision-making problems that are too complex to be solved atomically. If all sub-node decisions are routine to the degree that they all can be handled by a machine (i.e. no human-in-the-loop), we then speak of automation.
Optimization vs Decision-Making
Decision science is essentially an optimization problem. However, there is an important distinction between one-shot decisions
and sequential decision-making, the first being a (single) decision to be made only once (i.e. model optimization) and the latter being a real-time process
where decisions are made over successive time points (i.e. actions optimization).
Opportunistic vs Mission-Critical Decision-Making
For opportunistic problems
the performance, safety and risks are measured according to acceptable statistics (e.g. a 2 percent failure rate of a medical diagnosis, or a 40 percent failure rate of a financial trading algorithm) - their design usually is probabilistic
(i.e. deliberately exhibiting statistical variation). For mission-critical problems, however, failure is catastrophic (i.e. loss of mission) and tolerances are typically extremely small (e.g. railway and aviation traffic control systems, moon landings, ...) - their designs must be completely deterministic.