What are the different versions of decision theory?
The three main classes of decision theory are causal decision theory, evidential decision theory, and logical decision theory.
Causal decision theory (CDT) reasons about the causal relationship between the decision and its consequences. An agent using CDT selects the action that will physically cause the best expected outcome.
Evidential decision theory (EDT) reasons about the conditional probability of events given different choices. An agent using EDT selects the action it would be “happiest”[1] to learn it had taken. It views its action as one more fact about the world that it can reason about, and does not distinguish the causal effects of its actions from any other implications of taking them.
Logical decision theory (LDT) is a class of decision theories, including updateless decision theory, functional decision theory, and timeless decision theory, that use logical counterfactuals. An agent using an LDT acts as if it controls the logical output of its own decision algorithm, and not just its immediate action. LDTs match or outperform other forms of decision theory in problems such as Parfit's hitchhiker, the smoking lesion problem, and Newcomb's problem.
An example of an LDT is functional decision theory (FDT). FDT treats an agent’s decision as the output of a fixed mathematical function, and picks based on which output it would be best for this function to have, taking into account not just the consequences of the agent’s decision, but also all the other places the function is instantiated.
Further reading:
“Happiest” not in terms of its mood, but in terms of whether it expected the world to rank highly in its preference ordering ↩︎