READOS / Technology
Decision-Making Machines
How automated systems reason under uncertainty, plan ahead and learn from rewards, and how to test them before trusting them.
Start reading →Contents
- 01Decision making under uncertainty: agents, observations and actionsA plain-English introduction to how people and machines make decisions when outcomes, models, situations and other actors are uncertain, and the main ways to design decision systems.
- 02Bayesian networks: mapping how causes and evidence connectBayesian networks represent uncertain relationships as a map of influences. How they work, how they support diagnosis and prediction, and what they can and cannot tell you about cause and effect.
- 03Acting when you can't see everything: beliefs, filters and partial observabilityMany decisions must be made without knowing the true situation. How beliefs, Bayesian filters, Kalman and particle filters and belief-state planning handle partial observability.
- 04Sequential decisions: planning when today's choice shapes tomorrowMany decisions are steps in a sequence. A plain-English guide to Markov decision processes, policies, value functions, discounting and dynamic programming, with business examples.
- 05Looking ahead: tree search, rollouts and planning by simulationWhen problems are too large to solve completely, decision makers plan from the current situation by looking a few steps ahead. How lookahead, rollouts, Monte Carlo tree search and re-planning work.
- 06Reinforcement learning explained: learning good decisions from experienceReinforcement learning lets systems learn decision strategies by trial and error, guided by rewards. A plain-English guide to how it works, its main methods, its risks and where it fits in business.
- 07Learning from experts: imitation learning and its limitsSystems can learn by copying expert demonstrations. How behavioural cloning, cascading errors, expert correction and inverse reinforcement learning work, and what they teach about know-how.
- 08Be careful what you reward: designing objectives for decisionsAutomated systems and people both optimise what they are measured on. How objectives go wrong, how to balance multiple goals with trade-off analysis, and how to design rewards and KPIs that work.
- 09Testing decision systems before you trust themHow to validate automated decision systems: performance metrics, rare-event simulation, robustness to model errors, trade-off analysis, adversarial testing and responsible deployment.
- 10Tests that cannot fail: checking whether your assurance could ever find a problemA check that passes whether or not the problem exists tells you nothing, yet feels like assurance. How to spot tests that cannot fail, from padlocks to AI bias audits, and replace them.
These topic-based reading collections are organised alphabetically; chapters may be read in any order.
← All books