Book Radical Uncertainty provides a myriad of anecdotes about how “Bayesian reasoning” could be inadequate and derailed and deliberation and decision-making based on narratives (stories, scenarios, or, how the authors repeat throughout the book, “figuring out what is going on here”) is the state of the art in business, military, and other high-stakes decision-making and strategy. Then, a good strategy is such that is robust/prepared to at least two competing and sufficiently different scenarios/narratives of what is going to happen in the future.
Once the “reference narrative” (i.e., the most likely one) is chosen, agents should additionally “protect” it against risks that could jeopardise this narrative in particular. The authors criticise risk mapping (management practice) as being formulaic and reaching the core objective of “protecting the reference narrative”. I don’t quite understand this criticism – perhaps, this is a criticism of the application of the practice rather than the practice itself. Also, I guess the authors could direct the same criticism at FMEA.
Overall, debating plausible scenarios (the authors emphasize that it’s hard or impossible for a single person to generate multiple very different scenarios in an unbiased way and to argue for and against them at all sides, which is reflected in the management wisdom of Drucker, Sloan, Eisenhower, Bezos, etc.) looks just like sampling scenarios MCTS-style with GFlowNet or an RL model. Then, choosing the strategy that is robust to multiple plausible scenarios is just EFE optimisation.
And FWIW the plausible scenarios in this sampling could be Bayesian in the space of massive multi-factor distribution over the world states over the entire horizon of planning. However, this takes away another dimension of deliberation and decision-making: in group discussion, the decision-maker recognises that the contexts of the debaters don’t perfectly match and therefore the probabilities of the scenarios the debaters advocate for couldn’t be compared directly. The decision-maker not only weights the arguments but also decides whose reference frame/context to align with, which is a non-probabilistic decision by definition.