Back to LessWrong

Difference between revisions of "Updateless decision theory"

From Lesswrongwiki

Jump to: navigation, search
(writing stub)
(Content)
Line 11: Line 11:
 
to actions--across a probability distribution of all world-histories. ("Best" here, as in other decision theories, means one that maximizes a utility/reward function.)
 
to actions--across a probability distribution of all world-histories. ("Best" here, as in other decision theories, means one that maximizes a utility/reward function.)
  
This definition may seem trivial, but in contrast, CDT says that an agent should choose the best option at any given moment based on the effects of that action.
+
This definition may seem trivial, but in contrast, CDT says that an agent should choose the best option at any given moment based on the effects of that action. As in [http://lesswrong.com/lw/emc/causality_a_chapter_by_chapter_review/ Judea Pearl's definition of causality], CDT "cuts" the causal links inbound to the decider, treating this agent as as an uncaused cause. The agent is unconcerned about what evidence its decision may provide about the agent's own mental makeup--evidence which may suggest that the agent will make suboptimal decisions in other cases.
 +
 
 +
Evidential Decision Theory is the other leading decision theory today. It says that the agent should make the choice for which the  expected utility, as calculated with Bayes' theorem, is the highest. EDT avoids CDT's pitfalls, but also ignores the distinction between causation and correlation. EDT's insight is reflect in "UDT 1.1" a variant of UDT in which the agent takes into account that some of its algorithm (mapping from observations to actions) may be prespecified and outside its control, so that it has to gather evidence  and draw conclusions about this prespecified part of its own mental makeup.
  
 
==Blog posts==
 
==Blog posts==

Revision as of 06:35, 25 February 2016

Motivation

Updateless Decision Theory (UDT) is a new decision theory meant to deal with a fundamental problem in the existing decision theories: Treating the agent as a part of the world. In contrast, the most common decision theory today, Causal Decision Theory (CDT), the decision is not part of the world model--it is the output of the CDT, but in the context of the world the agent's decision is "magic": It is uncaused, like a dualist version of a soul with free will.

Getting this issue right is critical in building a self-improving artificial general intelligence. Such an AI must analyze its own behavior and that of a next generation that it may build.

Content

UDT specifies that the optimal agent is the one with the best algorithm--the best mapping from observations to actions--across a probability distribution of all world-histories. ("Best" here, as in other decision theories, means one that maximizes a utility/reward function.)

This definition may seem trivial, but in contrast, CDT says that an agent should choose the best option at any given moment based on the effects of that action. As in Judea Pearl's definition of causality, CDT "cuts" the causal links inbound to the decider, treating this agent as as an uncaused cause. The agent is unconcerned about what evidence its decision may provide about the agent's own mental makeup--evidence which may suggest that the agent will make suboptimal decisions in other cases.

Evidential Decision Theory is the other leading decision theory today. It says that the agent should make the choice for which the expected utility, as calculated with Bayes' theorem, is the highest. EDT avoids CDT's pitfalls, but also ignores the distinction between causation and correlation. EDT's insight is reflect in "UDT 1.1" a variant of UDT in which the agent takes into account that some of its algorithm (mapping from observations to actions) may be prespecified and outside its control, so that it has to gather evidence and draw conclusions about this prespecified part of its own mental makeup.

Blog posts

Relevant Comments

In addition to whole posts on UDT, there are also a number of comments which contain important information, often on less relevant posts.

External links

Problem Class Dominance in Predictive Dilemmas, section 3.4. (The best summary to date.)


See also