Notes on Harman's `Change in View'

David Traum (traum@cs.umd.edu)

My motivations:

 
Find suitable abstract account of human reasoning for implementation in artificial agent

• Practical, implementable
• close to common sense intuitions
• useful for language reasoning/interacting with humans

Harman Basics

 

• theoretical (beliefs) vs. practical (intentions, plans) reasoning
• reasoning vs. argument/proof

  reasoning

  actual changes, non-monotonic, inductive, non-logical

  argument

  principles of implication, monotonic information increase, deductive,

  no need for non-monotonic or inductive logic, or deductive principles of revision

• phases in reasoning:
  1. reflection (consider possibilities)
  2. revision (actual changes)
     not necessarily conscious
• Explicit vs. implicit belief

Logic and Reasoning

: defeasible revision principles

1. *Logical Implication Principle: The fact that one's view logically implies P can be a reason to accept P.
2. *Logical Inconsistency Principle Logical Inconsistency is to be avoided.
3. *(revision of 1) Logical Closure Principle One's beliefs should be closed under logical implication In other words, there is something wrong with one's beliefs if there is a proposition logically implied by them which one does not already believe in. In that case, one should either add the implied proposition to one's beliefs or give up one of the implying beliefs.
4. clutter avoidance One should not clutter one's mind with trivialities (explicit beliefs). Accept a new belief P, only if one has an interest in whether it is true,
5. *(revision of 1) recognized logical implication principle One has a reason to believe P if one recognizes that P is logically implied by one's view.
6. *(revision of 2) recognized logical inconsistency principle One has a reason to avoid believing things one recognized to be logically inconsistent.
7. (revision of 5) recognized implication principle One has a reason to believe P if one recognizes that P is implied by one's view.
8. (revision of 6) recognized inconsistency principle One has a reason to avoid believing things one recognized to be inconsistent.

Positive versus Negative Undermining

 

• Foundations theory of belief revision
  • keep track of reasons for belief
  • drop beliefs that have no satisfactory justification
  • add beliefs that don't need justification, or or are justified by other justified beliefs
• coherence theory of belief revision
  • don't need to keep track of reasons
  • don't need to justify ongoing beliefs
  • need justification only when there is reason to doubt
  • drop beliefs when incoherent
  • strive for minimal change to increase overall coherence
• Karen's aptitude test: (1) test results say scored low in history (2) believe low aptitude in history (3) believe history course scored well in was easy (4) decide not to take more history (5) find out score was erroneous.
  • foundations theory says drop all beliefs, reconsider all decisions based on (1). Seems normatively correct
  • coherence theory says keep beliefs about ease of course, etc., because coherent with other beliefs - only get rid of aptitude belief which is not consistent with new info. Seems descriptively correct, based on , e.g., debriefings of psych experiments
• Principle of Positive Undermining One should stop believing P whenever one positively believes one's reasons for believing P are no good.
• *Principle of Negative undermining One should stop believing P whenever one does not associate one's belief in P with an adequate justification (either intrinsic or extrinsic)
• Positive undermining (or drop reasons for believing P is no good). Problem with remembering justification and negative undermining is clutter avoidance
• Principle of Conservatism One is justified in continuing fully to accept something in the absence of a special reason not to.
• Full Acceptance  

  allows oneself to use P as part of starting point for future theoretical and practical reasoning

  stop investigating P

  -

  accepting as working hypothesis has only first principle

Principles of Belief Revision

 

1. clutter avoidance (see above)
2. Interest condition (on theoretical reasoning) One is to add a new proposition P to one's beliefs only if one is interested in whether P is true (and it is otherwise reasonable for one to believe P).
3. Interest in the Environment One has a reason to be interested in objects and events in one's immediate environment.
4. Interest in facilitating Practical reasoning If one desires E and believes M's being true would facilitate or hinder E, one has a reason to be interested in whether M is true.
5. Interest in Facilitating Theoretical Reasoning If one is interested in whether P is true and has reason to believe knowing whether Q is true would facilitate knowing whether P is true, one has a reason to be interested in whether Q is true (DRT: backward chaining?)
6. Interest in Not being Inconsistent

   tendency to avoid holding beliefs that are immediately inconsistent

   tendency to avoid holding beliefs one recognizes to be indirectly or less obviously inconsistent.

   Get Back Principle One should not give up a belief that one can easily (and rationally) get right back (so can't just drop the belief that a set of facts are inconsistent).

revising intentions

(practical reasoning)

• Principles of intention revision:
  1. principle of conservatism - minimize changes in intentions
  2. principle of coherence change intentions to to make them more coherent with each other and one's beliefs, discourage changes which make them less coherent
  3. principle of desire satisfaction encourage change that promise to promote the realization of one's ends, and discourage changes that promise frustration of those ends
• types of intention
  1. explicit - factors in further reasoning
  2. implicit?
  3. positive - involves the thought that it will lead to intended action
  4. negative - intention not to do something (which one might not do anyway, so doesn't cause lack of action, but this settles the matter)
  5. conditional - intermediate case - leads one to do something only if conditions occur.
• positive intentions require a plan - need to know a way to bring this intention about
• distinction between intentional action and action that is intended
• distinction between ``intends'' and ``aims'' (goal?) - e.g., aiming for a hole in one
• negative intentions are not aims
• rejection of putative principle that prediction excludes intention: ``One cannot for the intention of doing something one can predict one will do whether one has that intention or not.''

Reference

Gilbert Harman. Change in View. MIT Press, 1989.