tripleC 9(2): 494-501, 2011
ISSN 1726-670X
http://www.triple-c.at
CC: Creative Commons License, 2011.
Abduction as Incomplete Parameter Estimation
Moto Kamiura
kamiura@rs.noda.tus.ac.jp, moto@goo.jp, Faculty of Science and Technology, Tokyo University of Science,
Japan
Abstract: Abduction is a kind of logical inference, and has been studied in computer science and artificial intelligence (Fin-
lay and Dix 1996). Recently, Sawa and Gunji (2010) introduced a diagram to represent three types of inference: i.e. deduc-
tion, induction, and abduction, which are articulated by C.S.Peirce. Sawa-Gunji’s representation provides a new approach to
a numerical aspect of abduction. In the present paper, we show that Sawa-Gunji's representation of abduction is consistent
with Finlay-Dix's one, and integrate the two representations. Both parameter estimation and abduction occupy a similar
position on the integrated representation, although they are not completely corresponding. We present "incomplete" pa-
rameter estimation as a sort of "simulated abduction", which is a numerical aspect of abduction. It is applied to a first-order
autoregressive (AR(1)) model. As a result of numerical analyses on AR(1), the incompletely estimated parameter (IEP)
follows a Cauchy distribution, which has a power law of the slope -2 in the tail, although conventionally estimated parameter
is normally distributed. It is shown that the Cauchy distribution of the IEP is based on structure of ratio distribution of normal
random variables generated from the AR(1). This research suggests that the distribution of the IEP is not based on a mech-
anism of system itself, but on relationship between data structure on the given system (i.e. the given AR(1) process) and
one on the system observer (i.e. the estimator of the AR(1) parameter).
Keywords: Autoregressive model, Parameter Estimation, Observation, Cauchy distribution, Power law
Acknowledgement: Part of this work was carried out under the Cooperative Research Project Program of the Research
Institute of Electrical Communication, Tohoku University.
1. Introduction
C.S. Peirce who is a philosopher/logician classifies reasoning/inference into three types: i.e. deduc-
tion, induction and abduction (Peirce 1868; 1955). Abduction has been studied not only in philoso-
phy but also in artificial intelligence and computer science (Bylander et al 1991; Tanner and Jo-
sephson 1996; Finlay and Dix, 1996; Abe 2003). Recently, Sawa and Gunji presented a model
representing dynamic change in logical inference (Sawa and Gunji 2010). To define the model
dynamics, they propose an arrow diagram which is a tool to illustrate each of the three types of
inference. Sawa-Gunji’s representation of abduction is an operation in which an arrow in the dia-
gram is derived from the other arrows. This formalization for abduction is not only consistent with
the preceding studies (Peirce 1868; Finlay and Dix 1996), but also provides a new approach to a
numerical aspect of abduction (Kamiura 2010).
In the present paper, inheriting the Sawa-Gunji’s representation, we formalize a numerical as-
pect of abduction and find “incomplete” parameter estimation as a sort of “simulated abduction”.
Auto-regressive (AR(1)) model is used for numerical experiments and the results are shown. In the
conclusion, we discuss abduction mediating a relationship between system observer and power
law distributions.
2. Diagrams of Inference and Positioning of Abduction
Peirce assumes the three types of inference as the following operations: i.e. (i)Deduction: the major
premise and minor premise derive the conclusion. (ii)Induction: the minor premise and conclusion
derive the major premise. (iii)Abduction: the major premise and conclusion derive the minor prem-
ise.
tripleC 9(2): 494-501, 2011 495
CC: Creative Commons License, 2011.
Abduction includes an intrinsic incompleteness, which means that abduction is formally equiva-
lent to “the logical fallacy affirming the consequent”. Therefore, the context needs to be clearly de-
scribed to avoid unnecessarily confusing.
Let us recall a representation of abduction by Finlay and Dix (1996). It has almost the same
meaning which is given by Peirce (1955). Finlay and Dix use the representation of the first-order
predicate logic,
, (1)
where a and b are predicates and x is a variable. In standard logic, given a proposition a(x
0
) with a
value x
0
, one can deduce b(x
0
) (i.e. deduction). In abduction, one deduces the inverse: i.e. if one
knows b(x
0
), then one deduces a(x
0
). This is not permitted in standard logic, but is a formalization of
abduction.
Recently, Sawa and Gunji (2010) introduced the following arrow diagram, Figure 1, to represent
the three types of inference. Each inference is expressed as an operation in which one arrow is
derived from the other two arrows on this diagram.
Figure 1: Sawa-Gunji’s representation for
inference.
Finlay-Dix’s representation and Sawa-Gunji’s one are consistent with each other. Let us express
arrows of the Sawa-Gunji’s diagram by a := Minor premise, f := Major premise and b: = Conclusion.
When the arrow diagram is a commutative one and the dot on the lower left corner of the diagram
is considered as a set of the variable x, we can obtain , which is equal to (1). The
integration of Finlay-Dix’s representation and Sawa-Gunji’s one is expressed in Figure 2.
We can also confirm an example of conventional syllogism by the following substitution: a(x) = “x
is a human”, f = “If x is a human, then x is mortal” (i.e. “humans are mortal”) and b(x) = “x is mortal”.
Figure 2: Integration of the Finlay-Dix’s
representation and the Sawa-Gunji’s one.
In addition, we can see that abduction is accounted for by deriving the upper left arrow (i.e. “the
minor premise” or the map a) on Figure 1 or 2.
3. Transplanting Inference Structure to Numerical Function System
3.1. Modifying the Sawa-Gunji’s Representation
In this section, the diagram for logical inference in Figure 2 is converted into a diagram for numeri-
cal calculations. The following restriction is proposed: i.e. restrict the map a in Figure 2 to a map (
α
,
Minor
premise
Major
premise
Conclusion
a
a(x)
f
b
x
b(x)
496 Moto Kamiura
CC: Creative Commons License, 2011.
-), where
α
is a parameter and the hyphen – is a blank space for which a value x is substituted.
This procedure generates a diagram in Figure 3.
Figure 3: The diagram to transplant the structure of inference to numerical function system.
Each map in Figure 3 is interpreted as the following: (A1) the map adds infor-
mation of the premise (i.e. the parameter value
α
) of system structure f to the input data x. (A2) the
system transforms the input data x into the output data y, under the parameter
α
.
(A3) Composition of the maps (
α
, -) and f derives a set of the pairs of the input and output data,
. (2)
(A1), (A2) and (A3) respectively correspond to the minor premise, major premise and the conclu-
sion.
3.2. Relationship between Abduction and Parameter Estimation
In Finlay-Dix’s representation, abduction is an operation to derive the map a. The map
in Figure 2 is restricted to the map in Figure 3. The map (
α
, -) is
uniquely determined by the parameter
α
. Consequently, abduction can be regarded as deriving the
parameter value
α
from the two remaining arrows, f and b.
Based on this procedure, abduction is associated with parameter estimation in a context of con-
ventional science.
However, abduction includes an intrinsic incompleteness, which means that abduction is for-
mally equivalent to the logical fallacy affirming the consequent. In this point, abduction is different
from conventional parameter estimation. Therefore the study on a numerical aspect of abduction
advances to parameter estimation which includes some kind of incompleteness.
3.3. Abduction as Incomplete Parameter Estimation
Parameter estimation is based on two conditions: (B1) the model f reflects the data b
enough. (B2)
the adequate number of the data are used to estimate the parameter
α
.
(B1) and (B2)
Conventional Parame-
ter Estimation
(NB1) and (B2)
Model-Data (M-D)
Inconsistency
Table 1: Combination of the conditions on parameter estimation
(
α
, x)
f
b
x
(
α
, - )
f (
α
, x)
= y
= b (x)
tripleC 9(2): 494-501, 2011 497
CC: Creative Commons License, 2011.
Incomplete parameter estimation is defined by estimation which does not meet these conditions.
That is, in the incomplete parameter estimation, (NB1) the model f does not reflect the data
b
enough, or (NB2) a small number of data are used to estimate the parameter
α
.
The combination of the conditions on parameter estimation is organized as Table 1.
In this study, we consider only (B1) and (NB2). In the following section, using auto-regressive
(AR) model, we specify incomplete parameter estimation under the conditions.
4. Model Description
4.1. Conventional parameter estimation
Let us start from least-square method for parameters of AR(p) process. Assume that time series
is modeled by AR(p) process, , where and
, and ~ . The parameter is estimated from the first-order
condition on residual sum of squares (RSS) r
2
,
(3)
where
(4)
is linear predictor for the AR(p) process. If is a regular matrix, then estimated parameter
is expressed by
, (5)
where and .
4.2. Incomplete Parameter Estimation
In the context of least-square method for AR(p) parameters, the condition (NB2) presented in Sec-
tion 3.3, is specified by the following: (NB2) Data size of RSS, N, is small, i.e. , al-
though conventionally . If , then is a square matrix of . Under
this condition, we obtain an incompletely estimated parameter,
, (6)
where in .
5. Numerical Analysis
5.1. Distribution of Incompletely Estimated Parameter
Assume AR(1) time series, . Note that incompletely estimated parameter (IEP)
is different from the AR(1) parameter , which is given values.
From the AR(1), we obtain
, (7)
since . Now ~ , therefore ~ . From Eq. (6), we can obtain the
IEP as the following,
498 Moto Kamiura
CC: Creative Commons License, 2011.
. (8)
therefore
. (9)
On the other hand, From Eq. (7), we obtain . Thus, ~
. Generally, given two normal random variables, ~ and ~ ,
we obtain a random variable which follows a Cauchy distribution,
, (10)
where , and is the correlation coefficient between and . Conse-
quently, we can see that follows the Cauchy distribution,
, (11)
since and , where is an autocorrelation on time lag 1.
5.2. Peak Location of the Distribution and Robustness of the Distribution
Under the condition , we obtain
. (12)
Figure 4 shows autocorrelations of on the time lag 1, , for given AR(1) parameter
, where . Note that does not depend on the value of .
This graph shows that the location of the peak of , , is always negative although the
given AR(1) parameter .
Figure 4: The graph of autocorrelations of on the time lag 1, , for the given AR(1) pa-
rameter .
tripleC 9(2): 494-501, 2011 499
CC: Creative Commons License, 2011.
It is known that power law distributions generated by some nonlinear systems are fragile for pa-
rameter change (Kamiura and Gunji 2006). By contrast, the distribution (11), which shows power
law in the tail, is robust for the parameter change : i.e. If , then we obtain
and ~ , thus follows the distribution (11).
5.3. Computational Experiments
We can computationally confirm the analyses in the previous sections. Assume an AR(1),
, (13)
where ~ and .
Figure 5(a) and Figure 5(b) show a distribution of IEP given by Eq. (9), a normal distribution
( , ) and a Cauchy distribution given by Eq. (10)
( , ). Figure 5(b) is double logarithmic graphs.
Figure 5(a): The graphs of a distribution of , a normal distribution ( , ) and a
Cauchy distribution ( , ).
500 Moto Kamiura
CC: Creative Commons License, 2011.
Figure 5(b): The double logarithmic graphs of a distribution of , a normal distribution
( , ) and a Cauchy distribution ( , ).
Figure 6 shows asymptotic property of the distribution of defined by Eq. (5). The distributions of
get closer to a normal distribution with , as N gets larger. In this process, the power law
property of the Cauchy distribution in Figure 5 vanishes away. In N = 120, we can see conventional
parameter estimation, in which follows a normal distribution with and
.
Figure 6: The asymptotic property of the distribution of .
6. Conclusion
Abduction is a kind of logical inference, and has been studied not only in philosophy (Peirce 1868;
1955) but also in artificial intelligence and computer science (Bylander et al 1991; Tanner and Jo-
sephson 1996; Finlay and Dix, 1996; Abe 2003).
In this paper, we integrate Finlay-Dix's representation of abduction (1996) and Sawa-Gunji's one
(2010). Integrated representation derives a numerical aspect of abduction, which is identified as
incomplete parameter estimation. We introduce the incompleteness for parameter estimation of
AR(1) process.
As a result of the numerical analyses, the IEP follows a Cauchy distribution, which has a power
law of the slop -2 in the tail, although conventionally estimated parameter is normally distributed.
The Cauchy distribution is caused by a ratio distribution of normal random variables which is gen-
erated from the AR(1) process.
As one of the causes of power laws, self-organized criticality (SOC) on physical, chemical and
mathematical systems (Bak et al. 1987; Claycomb et al. 2004; Luque et al. 2008) has been known.
These power laws are based on the mechanisms of the systems themselves, which have some
organized interaction between their elements. On the other hand, power law which is derived from
the incomplete parameter estimation is not based on a mechanism of system itself but on relation-
ship between data structure on given system (i.e. given AR(1) process) and one on the system
observer (i.e. estimator of the AR(1) parameter). Consequently, our research suggests that com-
plexity expressed by a power law can be derived from the incomplete parameter estimation which
is a numerical aspect of abduction and is different from SOC mechanisms.
tripleC 9(2): 494-501, 2011 501
CC: Creative Commons License, 2011.
References
Awodey, S. (2006). Category Theory. Oxford: Oxford University Press.
Abe, A. (2003). The Role of Abduction in Chance Discovery. New Generation Computing 21(1), 66-71.
Bak, P., Tang, C., & Wiesenfeld, K. (1987). Self-organized criticality: An explanation of the 1/f noise. Phys. Rev. Lett., 59,
381-384.
Bylander, T., Allemang, D., Tanner M.C. & Josephson, J.R. (1991). The computational complexity of abduction. Artificial
Intelligence, 49, 25-60.
Claycomb, J.R., Nawarathna, D., Vajrala. V., & Miller, J.H. (2004). Power law behavior in chemical reactions. J. Chem.
Phys., 121(24), 12428-12430.
Finlay, J. & Dix, A. (1996). An Introduction to Artificial Intelligence. London: UCL Press.
Gunji Y.-P. & Kamiura M. (2004). Observational heterarchy enhancing active coupling. Physica D, 198, 74-105.
Kamiura, M. & Gunji,Y-P. (2006). Robust and Ubiquitous On-Off Intermittency in Active Coupling. Physica D, 218, 122-130.
Kamiura, M. (2010). Implication of Abduction: Complexity Without Organized Interaction. Proceedings of the 9
th
International
Conference on Computing Anticipatory Systems, accepted.
Luque, B., Miramontes, O. & Lacasa, L. (2008). Number theoretic example of scale-free topology inducing self-organized
criticality. Phys. Rev. Lett., 101, 158702.
MacLane, S. (1971). Categories for Working Mathematician. New York: Springer, 1997 (2nd ed.).
Marsaglia, G. (2006). Ratios of normal variables. Journal of Statistical Software, 16(4).
Peirce, C.S. (1868). Some Consequences of Four Incapacities, In: Charles Hartshorne and Paul Weiss, Collected papers of
Charls Sanders Peirce. Harvard: Harvard University Press, 1960.
Peirce, C.S. (1955). Abduction and Induction. In: Philosophical Writings of Peirce (Chapter 11), Dover Publications.
Sawa, K. & Gunji, Y-P. (2010). Dynamical Logic Driven by Classified Inferences Including Abduction. Conference Proceed-
ings of the 9
th
International Conference on Computing Anticipatory Systems, accepted.
Tanner, M.C. & Josephson, J.R. (1996). Conceptual analysis of abduction. In: J.R. Josephson and S.G. Josephson (Ed.),
Abductive Inference: Computation, Philosophy, Technology (Chapter 1). Cambridge: Cambridge University Press.
About the Author
Moto Kamiura
March 2007: Doctor of Science on Physics, Kobe University
June 2007 - March 2008 Technical Assistant in Kobe University
April 2008 - Present: Assistant Professor in Faculty of Science and Technology, Tokyo University of Science
April 2010 - Present: Cooperative Researcher in Research Institute of Electrical Communication, Tohoku University