Abductive Reasoning CS 621 – Artificial Intelligence Aarif Jindani(113050050) Alex Poovathingal(113050035) Ashok Rawat(113050018) Department of CSE IIT, Bombay 5 th Nov, 2011
Logical Reasoning Precondition ==Rule==> Conclusion Deduction – determine the conclusion Induction – determine the rule Abduction – determine the precondition ”Abduction is the source of all human knowledge”
History of Abductive Reasoning Charles Sanders Pierce (1839 -1914) Rule: All beans from this bag are white. Result: The beans are white. ”Abduction is no more nor less than guessing” Case: Therefore these beans are from this bag. (Add more)
Is Abductive Inference Correct? Rule: If an eclipse occurs, sky suddenly turns dark. Case: The sky is dark. Result: Eclipse occured. Rule: If it rained last night, the lawn will be wet. Case: The lawn is wet. Result: It rained last night. ”Induction and Abduction are fallible forms of reasoning. Their conclusions are susceptable to retraction. ” ”Success of our guesses far exceed that of random luck and seems born of attunement to nature by insticts. ”
Why is abductive inference useful? ”Abduction works often enough and is the only source for new ideas. ” When Newton saw the apple falling down, he must have done an abductive inference and came up with theory of gravity. A possible Thought Process Apple fell down. If earth pulled everything towards it, then ofcourse, apple too would fall down. So earth is pulling everything towards it.
Stages in Mental Process of Abduction ”Just as abduction originates with an emotional reaction, it ends with one. ”
Abductive Reasoning Process Model • Set-Cover Based • Defining a theory from a set of hypotheses based on the current observations. • Logic Based • Defining a logical theory based on a set of sentences (explanations) that describe the observations.
Set Cover Based Approach • A triplet (Φ, Ω, e) defines a domain of hypothesis assembly. • • Ω – Set of Observations • • Φ – Set of Hypotheses E – Mapping from subsets of Φ to subsets of Ω. Assumptions: • Computational For every subset Φ’ of Φ, e(Φ’) is computable. • Independence e(Φ 1 U Φ 2) = e(Φ 1) U e(Φ 2); for all Φ 1, Φ 2 that are subsets of Φ. • Monotonicity If Φ 1 is a subset of Φ 2, then e(Φ 1) is a subset of e(Φ 2). • Accountability α(φ) is the set of observations that cannot be explained without hypothesis φ.
Set Cover Based Approach • Abductive algorithm of Allemang, 4 parts: • Screening: Acceptability of all the possible hypotheses is decided and allocated in a hierarchical classification system. • Collection: Collection of hypotheses accounting for the observations. A set of hypotheses is made by adding every hypothesis that explains the observations. • Parsimony: Narrows down the collection to its most applicable subset. If a subset of the set of collected hypotheses is able to explain the observations, that is the new (narrowed down) hypothesis set. • Critique: Marks the most essential hypotheses among the available ones. Individually every hypothesis is excluded from the set and then the set is tested against the observations. If they cannot be proved, then the excluded hypothesis is marked essential.
Set Cover Based Approach Example: Consider a theory T consisting of the following propositions: (1) ∀x(bird(x) ∧ ~ ab(x) ﬤ flies(x)) (2) ∀x(ufo(x) ﬤ flies(x)) (3) ∀x(penguin(x) V ostrich(x) ﬤ ab(x)) (4) ∀x(songbird(x) ﬤ bird(x)) (5) ∀x(songbird(x) ﬤ eats insects(x)) (6) ∀x(frog(x) ﬤ eats_insects(x)) (7) ∀x(frog(x) ﬤ green(x) Ʌ croaks(x)) (8) ∀x(frog(x) ﬤ ab(x))
Set Cover Based Approach For the above theories we have the following domain of hypothesis assembly: Φ = {frog(x), songbird(x), ufo(x), no_bird(x)} Ω = {flies(x), green(x), croaks(x), ~flies(x), eats_insects(x)} e({ frog(x)}) = { eats_insects (x), ~ flies(x), green(x), croaks(x)} e({ songbird(x)}) = { eats_insects(x), flies(x)} e({ ufo(x), bird(x)}) = { flies(x)} e({ penguin(x)} ) = { ~flies(x)} e({ ostrich(x)}) = { ~flies(x)} e({ no_bird(x)}) = { ~flies(x)}
Set Cover Based Approach For a given set of observations: Ω’ = {~flies(F), croaks(F)} Result of Collection Phase: ~flies(F) => φ={ penguin(F) , ostrich(F), no_bird(F), frog(F)} croaks(F) => φ={ frog(F)} Hypotheses Set, HYP = { penguin(F), frog(F)} modify HYP to HYP = {no_bird(F), frog(F)} ---- incompatible Result of Parsimony Phase: e({frog(F)}) ~{ ﬤ flies(F), croax(F)} => HYP = {frog(F)} Result of Critique Phase: HYP = {frog(F)}
Set Cover Based Approach Limitations: • Basically the 4 assumptions • The computability of mapping ‘e’ of the subsets of hypotheses set to those of the observations in order to initiate the process. (Computational) • The independent assumption is quite strong, so restricting it to “easy to manage” domains only. • Changes in theory can lead to extensive respecification of the mapping ‘e’. • Restricted to diagnostic tasks or repair problems only as the mapping ‘e’ and causal relationship is known for them.
Logic Based Approach An abduction system consists of • a logical theory ‘T’ defined over the language ‘L’, and • a set of sentences A of ‘L’ that are called abducible. If a sentence φ is found as the result of an abductive process in searching for an explanation of ω, it must satisfy the following conditions: • T U φ is consistent, • T U φ ⊢ ω, • φ is abducible, i. e. , φ ε A.
Causal Logic Theory Konolige analyses: (C, E, T) is a simple causal theory defined over the firstorder language ‘L’, where C is a set of causes, E a set of effects and T is a logical theory defined over L. An explanation of a set of observations Ω subset of E is a finite set of sentences Φ such that: • Φ is consistent with T • T U Φ ⊢ Ω, where Ω is the conjunction of all ω ∈ Ω. • Φ is a subset-minimal.
Example Representation Example: The simple causal theory (C, E, T) is defined as follows: T is same as our example specified earlier. C = {frog(x), songbird(x), ostrich(x), penguin(x} E = {flies(x), green(x), croaks(x), eats insects(x)} If we have the set of observations Ω = {~flies(F), croaks(F)}, then Φ = {frog(F)} is an explanation because • frog(F) is consistent with T, • T U frog(F) ⊢ ~flies(F) and croaks(F), and • frog(F) is subset-minimal.
Abductive Logic Programming • Extension to logic programming with abduction • Separates theory in two parts – Normal Logic Program to identify Φ (Backward Reasoning) – Integrity Constraints to filter set of possible candidates
Example Abductive Logic Programming • Logic Program – Grass is wet if it rained. Grass is wet if the sprinkler was on. The sun was shining. • IC – False if it rained and the sun was shining.
Power of Abduction ”Abduction is not a feeble-minded cousin of deductive principles like modus ponens. Its actually, a far richer and more powerful form of thinking”. People's understanding of causality is inherently nonverbal because it is rooted in visual and kinesthetic perception.
Applications of Abduction in Real-Life Domains Targets to be Explanained Explanatory hypothesis Science Experimental Results Theories Medicine Symtoms Diseases Crime Evidence Culprits and Motives Machines Operations, Breakdowns Parts, Interactions and Flaws Social Behaviour Mental States and Traits ”Early men hypothesized the existence of God inorder to explain the design and existance of the world. ”
Applications in Computer and AI Fault Diagnosis Automated Planning Medical Reasoning
Conclusion Abduction is the qaulitative, everyday reasoning. Deduction shows that something must be, induction shows that something exists, and abduction shows that something mabye. Abduction is that starting point of all research.
References Thagard, P. (2007). Abductive inference: From philosophical analysis to neural mechanisms. Cambridge: Cambridge University Press. Gabriele Paul (1993). Approaches to abductive reasoning: An overview. German Research Center for Artificial Intelligence. Erkki Patokorpi: Role of abductive reasoning in digital interaction, 2006 Drew V. Mc. Dermott and Jon Doyle, Non-monotonic logic I: MIT AI Lab Memo 468 (1978). http: //commonsenseatheism. com/? p=3703 http: //en. wikipedia. org/wiki/Abductive_reasoning
Скачать презентацию Abductive Reasoning CS 621 Artificial Intelligence Aarif
3c48b1b1222c1c67f255a3f17e03ec00.ppt