INVESTIGATING THE SEMANTICS OF REMOTE PERCEPTION WITH SIMILARITY ESTIMATES AND MULTIDIMENSIONAL SCALING.

• Document Type: CREST
• Collection: STARGATE
• Document Number (FOIA) /ESDN (CREST): CIA-RDP96-00789R002200130001-3
• Release Decision: RIFPUB
• Original Classification: K
• Document Page Count: 30
• Document Creation Date: November 4, 2016
• Document Release Date: October 14, 1998
• Sequence Number: 1
• Content Type: BOOK

Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3 (~q
October 1987
INVESTIGATING THE SEMANTICS OF REMOTE.
PERCEPTION WITH SIMILARITY ESTIMATES AND
MULTIDIMENSIONAL SCALING.
By:
S. James P. Spottiswoode.
Prepared for:
Dr. Edwin C. May
Geosciences and Engineering Center.
SRI International,
333 Ravenswood Avenue,
Menlo Park,
California 94025.
9421 Charleville Boulevard,
Beverly Hills,
California 90212 U.S.A.
[phone]
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
October 1987
INVESTIGATING THE SEMANTICS OF REMOTE
PERCEPTION WITH SIMILARITY ESTIMATES AND
MULTIDIMENSIONAL SCALING.
By:
S. James P. Spottiswoode.
Prepared for:
Dr. Edwin C. May
Geosciences and Engineering Center.
SRI International,
333 Ravenswood Avenue,
Menlo Park,
California 94025.
9421 Charleville Boulevard,
Beverly Hills,
California 90212 U.S.A.
[phone]
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
ABSTRACT
In experimental studies of remote perception the analysis of the resulting data for accuracy and
information content has utilized techniques based on rankings by judges and encodings of targets and
responses with sets of descriptors. Geographical locations have frequently been used for targets, but
there have been few studies which investigate what attributes of such material are preferentially conveyed
or how such perceptions are represented. Elucidation of these questions might permit the construction of
improved methods of judging remote perception experiments. The analogous problems in normal
perception have been investigated by studying the structure of similarity measurements and is the method
here applied to the remote perception case. In this pilot study subjects estimated the global similarity
between pairs of photographs of geographic locations. The resulting matrix of similarity values was
analyzed by multidimensional scaling to give and two and three dimensional representations of the
psychological data. A method of estimating the deviation from chance expectation for such data is
developed. The results for the remote perception pilot study are compared with structures derived by
multidimensional scaling from a comparison study using the same targets normally perceived. The
remote perception study shows no deviation from chance by the criterion developed here but the resulting
two-dimensional semantic structure shows parallels with that from the comparison study and gives weak
evidence for the existence of the underlying semantic dimensions of predominately man-made scenes
versus predominately natural scenes and the presence versus the absence of land-water interfaces in the
scenes.
I
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
TABLE OF CONTENTS
ABSTRACT ........................................................................................................
LIST OF ILLUSTRATIONS ................................................................................... iii
LIST OF TABLES ................................................................................................ iv
1 INTRODUCTION ........................................................................................ 1
1.1 Similarity methods ............................................................................ 2
1.2 Multidimensional Scaling ................................................................... 2
1.3 An Example of Multidimensional Scaling of Geographical Data ................. 3
1.4 Criticisms of Similarity Studies in Environmental Psychology ................. 5
1.5 The Assessment of Mean Chance Expectation ........................................ 6
1.51 Mean Chance Expectation Calculation by Noise Addition .............. 7
1.52 Mean Chance Expectation Calculation by Permutation .................. 8
2 PILOT STUDY ............................................................................................. 11
2.1 Method ............................................................................................ 11
2.2 Subjects ............................................................................................ 12
2.3 Targets ............................................................................................ 12
2.4 Target Selection Method ...................................................................... 14
2.5 Results ............................................................................................. 15
3 COMPARISON STUDY ................................................................................. 19
3.1 Method ............................................................................................. 19
3.2 Results ............................................................................................. 19
4 CONCLUSION ............................................................................................. 20
5 APPENDICES
5.1 Appendix 1: Multidimensional Scaling Algorithm .................................. 21
5.2 Similarity Data from the Pilot Study ...................................................... 22
6 REFERENCES .............................................................................................. 23
II
Approved For Release 2000/08/08: CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
LIST OF ILLUSTRATIONS
1. Plot of 2 Dimensional MDS Solution from Comparison Study ................................ 4
2. Plot of 3 Dimensional MDS Solution from Comparison Study ................................ 5
3. Plot of Stress Against Treatment of Data for 10 Points ........................................... 8
4. Plot of Stress Against Treatment of Data for 13 Points ........................................... 9
5.
6.
7.
8.
9.
Scale for Subjects to Indicate Similarity Estimate .................................................
14
2 Dimensional MDS Structure from the Pilot Study for Both Subjects Combined .......
17
2 Dimensional MDS Structure from the Pilot Study for Subject AV
........................
17
2 Dimensional MDS Structure from the Pilot Study for Subject DK
........................
18
3 Dimensional MDS Structure from the Pilot Study for Both Subjects Combined ......
18
Approved For Release 2000/08/081 1 CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
1. Dissimilarity Estimates from Comparison Study ................................................. 4
2. Targets for Pilot Study .................................................................................... 13
3. Stress Values for MDS Models of the Pilot Study .........................:...................... 16
4. Stress Values for MDS Models of the Comparison Study ...................................... 20
Approved For Release 2000/08/08 '(CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
An extensive literature now exists on a form of communication in which persons acquire
significant information about some visual scene with which they have no sensory connection.2,3,11
Typically such studies will employ target material which is relatively complex and unconstrained in
comparison to the smalls sets of symbols like Zener cards used in many earlier parapsychological
experiments. Target material in these studies may consist of pictures of geographical locations while
the information produced by the subjects comprises more or less fleeting visual, tactile, kinesthetic and
other impressions. This phenomenon has suffered from some terminological confusion, being denoted
clairvoyance, remote viewing and remote perception by various workers; we shall use the latter term.
A key problem in the study of remote perception revolves around the comparison of the target
material with the subject's response. The methods used for the assessment of remote perception data
have progressed from matching and ranking by judges to methods based on descriptor sets.5 The
descriptor method has several advantages but has raised the question of whether judging methods could be
further improved by designing a descriptor set which was optimized for the semantic structures
preferentially available for apperception in remote perception experiments. It may be possible to base the
quantification of the information content of transcripts on a small set of underlying semantic dimensions
which might serve as "basis vectors" for the subject's internal representation of the target. If such basis
vectors could be found, complex constructs in the viewing data might be assembled by combining sets of
data so expressed.
Considerable work has been done on solving the analogous problems for visual
perception,3,4,10,16 and a variety of techniques have been utilized for assessing the semantic structures
created by a stimulus, with semantic differential scales, adjective lists and similarity estimates being
frequently employed. It is possible to extract considerable information from these data, which may be
broadly classified as objective, that is related primarily to the target material' and subjective, that is related
to the internal state of the subject. In particular, work has focussed on assessing the semantics of
response to geographic and architectural environments; the field has become known as environmental
psychology. In an important paper, Ward and Russell 16 compare several measurement techniques
across a set of common stimuli which were pictures of geographical locations. They showed that using
semantic differential scales and similarity estimates, it is possible to extract objective information about
targets in terms of a small number of semantic dimensions such as predominately natural scenes versus
predominately man-made scenes, open scenes versus enclosed scenes and the presence versus absence of
land-water interfaces in the scenes. Subjective information is also determinable and examples of this are
Approved For Release 2000/08/08: CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
the affective dimensions of pleasure versus displeasure and arousal versus boredom. The relative
subjective information rates of target scenes can also be estimated 10 as can information about preferred
behaviors in the presented environment. 16
1.1 Similarity Methods
Among the various methods employed in these studies, that based on estimates of similarity given
by subjects: enjoys two advantages. Firstly, there is no linguistic component in the subject's task and
therefore no assumptions have to be made about whether the words used in the semantic differential
method, for instance, are relevant to the descriptive task or have a constant definition. In the case of
remote perception data this might be a substantial advantage since little is known of what
representational structures are employed. The second benefit of similarity information is that it exhibits
less variance than semantic differential or adjectival methods 16 when the data is reduced to a small
number of semantic dimensions.
In a typical study employing the similarity method subjects are presented with all possible pairs
of targets drawn from a set of targets. They are required to make a global similarity estimate of each pair
of targets on a scale from 1 (very, very similar) to 7 (very, very different). The ordering of presentation
is counterbalanced. The resulting symmetric matrices of similarity estimates can be analyzed for
intersubject reliability and for a main effect due to the target pairs. Further analysis has used
multidimensional scaling (MDS), usually with a Euclidean metric, to represent the similarity data in a
space of chosen dimensionality. As a typical example, slides of 20 locations were used as stimuli. 16
The resulting 190 pairs of targets were assessed by 41 subjects. Intersubject reliability was
approximately 0.9, a main effect in the data due to the stimulus pairs was found at a level of p < 0.0001.
MDS analysis in a 5 dimensioned Euclidean space, followed by rotation of the structure, yielded the
following dimensions: natural, water scenes versus man-made, land scenes; natural scenes versus man-
made scenes; open versus enclosed scenes; natural, open scenes versus man-made enclosed scenes; and
natural water scenes versus natural land scenes.
1.2 Multidimensional Scaling
MDS is a technique in which similarity, or dissimilarity data, can be represented in a metric space
of chosen dimensionality. The method finds a configuration of points in the given space, one for each of
the stimuli, such that stimuli which are adjudged similar are close and ones found different are far apart.
Formally, the technique finds a set of points in the given space such that the distances between them are
maximally monotonic with the dissimilarity data. The resulting structure can be rotated so as to find
dimensions with easily interpretable semantic content. The method takes as input a matrix of
Approved For Release 2000/08/01: CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
dissimilarities, dij, and attempts to find n points such that the distances between the points, Dij, in the
chosen space are such that,
(dij Dij )
ii j
ix)
is minimized. S is the stress index defined by Kruskal.6'7 Thus the Dij are as much alike the dij as
possible in a least squares sense. This procedure can be run in a space of any dimensionality for the Dij,
Many algorithms have been developed for this purpose; the one used in this study follows that due to
Kruskal. This nonmetric MDS algorithm assumes that measurements are nominal or ordinal in
character. It is therefore appropriate for the dissimilarity estimates and it has been widely used for this
purpose in psychology and sociology. It may be mentioned that the nonmetric algorithm employed in
this study produces solutions which are invariant under any monotonic transformation of the input
dissimilarity data.
MDS bears some resemblance to factor analysis and principal components analysis. However,
MDS is capable of extracting fewer salient dimensions from a set of dissimilarity data than either of these
techniques. A further difference is that the two above-mentioned techniques assume that there is a linear
relationship between the variables and stimuli and, because of their inherent linearity, these methods tend
to overestimate the number of dimensions needed to satisfactorily represent some data. For further
details of the algorithm used in the pilot study see Appendix 1.
1.3 An Example of the MDS Analysis of Geographical Data
As an example we take the comparison study described below wherein the 10 targets used in the
remote perception pilot study were assessed for similarity by normal perception. Three subjects were
presented with all 45 possible pairings of the 10 targets in a randomized order of presentation. Each
subject marked the global similarity of the pair of targets on a scale going from very, very similar to
very, very different. The estimates from the 3 subjects were averaged to give the following dissimilarity
matrix:
Approved For Release 2000/08/08 CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
Table 1
DISSIMILARITY ESTIMATES FOR COMPARISON STUDY
Target Number
Plots of the 2 and 3 dimensional MDS solutions for this data are given below.
Waterfall
Valley Mountain
? .
.Fields
Grand Canyon
0 j
Port
?
City with canal
? City
FIGURE 1. PLOT OF 2 DIMENSIONAL MDS SOLUTION FROM COMPARISON STUDY.
Approved For Release 2000/08/084: CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
In this 2 dimensional solution the Y. axis appears to be interpretable as predominately natural
versus man-made scenes and the X axis as predominately horizontal versus vertical elements in the
scenes.
Grand Canyon
?
T Waterfall
Dune coastline
Fields
if lLake
T mountain
City
?
Por t*
City with canal
?1
FIGURE 2. PLOT OF 3 DIMENSIONAL MDS SOLUTION FROM COMPARISON STUDY.
This 3 dimensional solution further separates and clarifies the semantic components. Reading
from negative to positive, the X axis gives man-made versus natural scenes, the Y axis gives no land-
water interface versus land-water interface while the Z axis. goes from primarily horizontal images to
scenes with strong vertical components. These interpretations are speculative as the data set is very
small, but they may serve to demonstrate the method.
1.4 Criticisms of MDS Studies in Environmental Psychology
Tversky has pointed out that it may not be possible to map similarity data into a metric space by
MDS or similar methods because psychological similarity data may be incompatible with the
assumptions of such a representation. 14,15 He cites the example of similarity estimates of countries;
5
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
for instance Cuba might be ranked as globally similar to Russia and globally similar to Jamaica.
However Jamaica and Russia would not be estimated as globally similar. Thus the triangle inequality
would be violated and therefore no model based on metric spaces can represent this kind of data
adequately. Such concept crossing semantic structures may contribute stress to the models presented here.
A further problem with similarity data is that examples are easily found which show that similarity is not
a symmetric relation. Writing aSb for a is similar to b, that is, aSb #> bSa. Thus distance metrics,
for which D(a,b) = D(b,a), will again not adequately represent such data.
Another criticism of environmental psychological models of the kind considered here was raised by
Daniel and Ittelson.1 The structures resulting from MDS of similarity estimates obtained from subjects
viewing photographs of locations may be approximately reproduced by replacing the photographic
stimuli with terse verbal descriptions of those locations. Little difference to the MDS structure results
if subjects are presented with the two statements "Kansas wheat field" and "sunset over a mountain lake"
and asked to rate their "global dissimilarity" or whether they are shown pictures of these scenes and asked
the same question. It is argued that whenever subjects are required to respond to very diverse
environmental settings with unfamiliar, poorly specified, or inappropriate response categories, little or
nothing will be learned about the effects of features of the environment on perception or behavior.
Instead, there will be a tendency for a priori semantic relationships to emerge, relationships that are only
very abstractly related to the physical features of the environment. However, the similarity method does
reliably discern the differing targets. For instance target 4 might be far from target 9 in the NMS
structure and this might be true whether targets 4 and 9 are presented as short phrases, pictures or visits to
the actual locations. The modality of the information obtained by a respondent in remote perception
protocols is very variable, a fact that plagues attempts to construct descriptor sets which cover specific
and abstract components at the same time. The insensitivity of the similarity approach to this difference
might therefore be advantageous in the remote perception context.
1.5 The Assessment of Mean Chance Expectation by Noise Addition and
Permutation
Similarity estimates are unusual as parapsychological data in that there is no "correct" answer for
the similarity estimate obtained in each trial. How then are such data to be assessed for extra chance
effects? An obvious, but inapplicable, method is to examine successive measurements of each dij for
any tendency to clump around one value. This method could not be used in the pilot study since many of
the dij would only have one experimental measurement. An alternative method of finding a chance
expectation requires an additional hypothesis. If it is assumed that the dissimilarity estimates obtained
from a remote perception protocol do in fact have a structure compatible with a low dimension Euclidean
space, then it is possible to compare the stress value of the MDS configuration for the similarity data
6
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
with stress values from random pseudo-similarity data. This is the method employed in the pilot study.
The additional assumption required for this method certainly holds for the type of targets used in the pilot
study when assessed for similarity by normal perception, as in the comparison study given here. Whether
it remains valid under remote perception conditions is unknown.
To implement this method it is necessary to examine the relationship between stress and the
standard deviation of the injected noise. This relationship is also a function of the number of points and
the dimension of the space of the calculation. These four parameters, stress, noise, dimensionality of
representation and number of points are connected by functions for which closed form solutions are not
known. There is also very little numerical work on them in the literature. 12,17 Two methods for
evaluating the mean chance expectation (MCE) for stress values from MDS's of dissimilarity data were
explored in this study. The first, examined by Young,17 involves the generation of a dissimilarity
matrix by the following method:
MCE Calculation by Noise Addition
(a) Choose coordinates of m points in a space of n dimensions. Let the
coordinates of point i be (xlj} u = 1,...,n).
(b) To calculate the distance between points i and j perturb the coordinates of
points i and j by adding a random normal deviate, R of mean = 0 and
chosen standard deviation, to all of the coordinates of each of the points.
Then calculate the distance dij from
R) - (xJk + R) )2 ) 0.5
d 1j _
(c) Repeat (b) till all the dij are found.
(d) Compute the stress for an MDS solution of the dij so obtained in an n
dimensional space.
(e) Repeat steps (b) through (d) till enough values of the stress have been
obtained for the distribution of stress values to be known.
Whilst the above method of injecting noise into a set of coordinates is useful for exploring
relationships between stress, noise level, number of points and number of dimensions, an alternative
method which gives an exact measure of the mean chance expectation in an experimental context is
provided by permutation.
7
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
.225
.2
.175
.15
.125
.1
.075
.05
.025
0
-.025
f _ f
0 0.05 0.1 0.2 0.3 Permuted
Standard Deviation of Noise
1.52 MCE Calculation by Permutation.
(a) Measure a set of experimental dissimilarity measurements dij,
(b) Compute a MDS configuration for this data in a space of chosen dimension.
Let the resulting stress be Se.
(c) Generate a random permutation of the dij. That is dij -> dkl where k,l are
random integers in where m = n(n - 1)/2.
(d) Compute an MDS structure for the permuted matrix. Call the stress for this
configuration Sp.
(e) Repeat c and d till enough values for Sp are obtained that the distribution of
Sp is well defined.
(f) Compare the experimental stress value Se with the distribution of Sp's and
compute the chance probability by direct integration or other methods.
These methods can be exemplified by taking the data from a normal perception control study of
the 10 locational targets used in the pilot study. The following plot shows the effect of adding normally
distributed noise to an exact array of 10 points, following the first of the above methods, and also the
stress values from permuted matrices from the comparison study in this report.
Approved For Release 2000/08/0% : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
As can be seen, the two methods give close agreement for the chance level stress value for this
number of points in a 2 dimensional representation. The same exercise can be repeated for other numbers
of points. For instance with 13 points in 2 dimensions:
.15 . ....................................................................... f..................................................... ....... ........................................
0 .05 0.1 0.2 0.3 Permuted
Standard Deviation of Noise
As can be seen from the above, it is possible to use the permutation method to distinguish extra
chance similarity values in the pilot study, or similar experiments, always providing that the additional
assumption of low stress structure in the noise free data is assumed. It can also be seen that as the
number of points in the configuration is increased the magnitude of the chance expectation stress
increases whilst its variance decreases. Thus for a given experimental stress value, the Z-score of the
stress value will be larger the greater the number of points. However, in designing a remote perception
study one would like to use the minimum number of points so as to reduce the necessary number of
trials. The number of trials is, of course, proportional to the square of the number of points. The
question therefore arises what minimum number of points is required to establish a structure. Young17,
continuing investigations by Shepard, shows that the accuracy of the metric data recovered from an MDS
of noisy data increases as the number of points increases, and varies inversely with the increase in error
and with the number of dimensions scaled. Young's data show that with 10 points a 2 dimensional
structure can be recovered with 72% accuracy of the original distances, or dissimilarities, in the presence
of a.normally distributed error of standard deviation one half of the standard deviation of the true distances
in the configuration. Metric recovery of 3 dimensional structure is 68% under these conditions which
require 45 similarity values to be estimated. A 10 target set was therefore chosen for the pilot study as a
9
Approved For Release 2000/08/08 :CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
compromise between the desire to increase the number of points allowing for good metric recovery and
tightly distributed MCE stress values and the requirement to keep the number of similarity values to be
measured small.
This analysis can be extended to the case where part of the similarity matrix is unmeasured and
where those elements lacking an experimental value are filled by the mean value of the matrix elements.
In this circumstance, investigation has shown that the variance of the stress values of the permuted
matrices is increased and thus a given low value of stress from experimental data will have a reduced Z-
score with respect to the broadened chance distribution. However experiments have shown that the
method is quite tolerant to missing data and provided that less than 30% of the matrix elements are lost,
the recovered 2 dimensional structure (with 10 data points) is usually close to the structure derived from
the full similarity matrix.
10
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
2 PILOT STUDY
The aim is to collect dissimilarity estimates for pairs of targets from a set of 10 photographs of
geographical locations in a remote perception protocol and analyze these by multidimensional scaling.
As this study is very much conceived as exploratory in nature there are no formal null hypotheses.
Similarity data has never before been obtained in this experimental setting and therefore it is unclear
what results may be produced. However, by analogy with the MDS studies of similarity data of
geographical stimuli discussed above, it may be expected that the MDS structure resulting from, the
similarity estimates obtained through remote perception will show a stress below that expected by
chance. The chance expectation value of the stress will be calculated by the permutation method described
above; the experimental values for dissimilarities will be randomly permuted and a chance expectation
distribution of stress values thereby found. The experimental stress will be compared to this distribution.
Two and three dimensional MDS analyses will be so performed.
2.1 Method
The set of 10 targets used in the pilot study require the measurement of 45 quantities for the matrix
of similarities. Owing to the necessity of keeping the subjects blind to the target pair used in each trial it
was necessary to choose the target pair for each trial by random selection from the whole set of 45 pairs.
Using the pool without replacement would otherwise have been preferable, since then 45 trials would
have guaranteed that each target pair had a trial assigned to it. For the utilized method of selection with
replacement, the planned 45 trials for each subject would give a fair chance of 80% coverage of the
similarity matrix for the combined subject data. Since we are trying to ascertain 45 continuous numerical
quantities with the very noisy remote perception channel, many more trials than this would be preferable.
To the author's knowledge, this experiment breaks new ground in remote perception protocols in
several ways and some aspects of it appear problematic. Firstly subjects are required to produce
descriptions of two targets in each trial before feedback is presented. The success of the method depends
on their ability to produce some accurate information to both targets in each trial; successful description
of only one will not suffice as there will be no information on which to base the similarity estimate.
Secondly, the subjects become familiar with all the 10 targets after a few trials and this complicates their
task because of their propensity to fixate on one of the targets from memory rather than the fleeting, and
often incomprehensible, imagery thought to constitute ideal remote perception data. Both these effects
were hopefully ameliorated by encouraging the subjects to limit the time of their efforts to an absolute
minimum and to produce extremely sparse transcripts of the two targets. Subjects were requested to cease
their remote perception-effort as soon as they felt they had some visual material and an overall gestalt of
11
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
the scene. Another concern was whether the subjects would be able to correctly assign the remote
perception information to each target in a trial. If some of the perceptual elements were to be mixed
between the targets, the pair might be judged over-similar. If there was a consistent bias of this kind
across the experiment, all the similariy estimates would be biased towards excessive similarity and the
resulting MDS structure would be unchanged, since the structure is invariant under monotonic
transformations of the experimental data. However, the effect might not be so consistent. Also it is
possible that the order of the transcripts might not correspond to the target order; in other words the first
transcript produced might refer primarily to the second target and vice versa. This should not effect the
similarity estimate and hence the MDS analysis.
2.2 Subjects.
Two subjects were chosen for this pilot study; each being male and in their thirties. One, AV, had
considerable experience of remote perception experiments. The other, DK, had not previously taken part
in any formal parapsychological studies, but was of the belief that he could successfully remotely
perceive targets. The purpose of the experiment in elucidating the semantics of remote perception was
explained to both participants and both appeared highly motivated to have the experiment succeed.
2.3 Targets.
In selecting ten scenes of geographic locations for the pilot study, several criteria were used:
(a) To use targets which were similar to those in use at SRI International so
that compatibility of the database would allow conclusions to be drawn from
the pilot study which would find application at SRI. It was thought that the
data from the pilot study might be analyzed by a descriptor set method and a
comparison drawn between the estimated similarities from the pilot study
and the encoded transcripts.
(b) To include in the set several pairs which would have been judged fairly
similar, and several very different, on the basis of Ward and Russell's 16
results. On the assumption that similarity estimates derived by remote
perception approximate those derived from perceptual studies, this would
ensure that the target set for the pilot study would have target pairs at the
extremes of the similarity - dissimilarity axis. In the low signal to noise
ratio regime of remote perception studies this maximizes the possibility of an
extra-chance result.
(c) To include targets at the extremes of the semantic dimensions found in
normal perception similarity studies of geographic locations.
12
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
Table 2.
TARGET POOL FOR PILOT STUDY.
Pilot Study
SRI Target
Target
Target Description
Target No.
Number
Short Name
1
-------
----------
Waterfall
-----------
- ---------------------
Waterfall, dark vertical rock, trees.
2
393
-
Mountain
--------------------------------
Vertical rock and snow fields. Dark sky.
-----------
3
------------
323
--------------
Mountain Lake
-----------------------------------
Blue crater lake, snow, clouds, small cliffs.
------------
4
-----------
414
-------------
Grand Canyon
---------------------------- ----
Many jagged pinnacles, all yellow and brown.
-----------
5
-----------
334
--------------
Valley
----------------------------------
Deep U shape, mountains, woods,snow,river.
6
Port
Harbor, city buildings, boats.
7
447
Dune coastline
--------------------------------
Wind sculpted dune field, slopes to beach, sea.
8
401
Fields
--------------------------------
Flat Kansas fields, rectilinear pattern, church.
9
334
City with canal
-------------------------------
Canal, city with skyscrapers, bridge.
-----------
10
------------
382
------------
City
-----------------------------------
Old European city, church spires, hills.
The pilot study targets include three city scenes, and six scenes where there is little evidence of
man-made activity. Also present are six scenes with water (either rivers or shoreline) clearly visible and
four with no water present. Thus the set includes several targets which would probably be adjudged
similar and several which are very dissimilar as well as targets which are at the extremes of the man-made
versus natural and land-water interface versus no land-water interface semantic dimensions. The ten
targets chosen for the study were identified from the large collection of pictures of geographical
13
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
locations at SRI. The pictures give a wide view over the scene and some of them appear to be side
looking aerial shots. All were reproduced onto color photographic prints of 20cm. x 25cm. size.
2.4 Target Selection Method
A program was written in FORTRAN to control the target selection in the experiment. This
program incorporated a pseudo random number generator algorithm due to Lewis & Payne.8 The output
of the implementation of their algorithm used in this experiment was checked against values given in the
cited reference to ensure correct functioning. The algorithm uses the feedback shift register method and
has been extensively checked for randomness.
The random number generator was seeded once at the beginning of the study prior to the first trial.
Thereafter all targets pairs were determined by this choice; the targets for a given trial being unknown to
the experimenter and subjects until that trial's completion. This method of target selection was adopted
in order to reduce any effects of intuitive data selection due to the experimenter.9 With the method used
here only one decision point occurred in the entire experiment (rather than one decision point for each
trial) and consequently the opportunity for fortuitous target selection by paranormal selection on the part
of the experimenter was largely eliminated.
In each trial the subjects were given the following instructions:
"You are required to do two precognitive remote viewings on two target pictures. These pictures
are of geographical locations somewhere around the world. The viewings will be rather quick and the aim
is to perceive a few of the most salient features of the locations. You are encouraged to make drawings
and verbal comments. It is suggested that you might describe the first target for approximately five
minutes, take a short break and then describe the second target for a similar period. You will then be
asked to mark your estimate of the similarity of the two targets on a scale as below:
Estimate similarity...
Uery, very similar
y
FIGURE 5. SCALE FOR SUBJECTS TO INDICATE SIMILARITY ESTIMATE.
different
Uery, ver
14
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
Immediately after this you will be shown the first of the two targets and allowed to examine it. A
minute later the second target will be shown. This completes the trial. It is hoped that we can complete
two, or perhaps three, trials per session."
Each trial took approximately 30 minutes and followed the above sequence. Subjects made
drawings with written comments for each target description and took a break of approximately 5 minutes
between remote perception efforts on the two targets. There was no delay between the end of the second
remote perception and the entering of the similarity estimate which was then followed by immediate
feedback to both targets. The position of the mark made on the scale by the subject was read off on a
separate scale in which "very, very different" corresponded to 10 and "very, very similar" to 0. The scale
was read to one decimal place.
2.5 Results
Owing to limitations of experimental time only 65 of the planned 90 trials were completed. One
subject, AV completed 35 trials giving values for 24 of the 45 matrix elements, while the other subject
DK completed 30 trials which also gave 24 measured matrix elements. In order to examine the data from
both subjects together, each subject's dissimilarity estimates were normalized to unity standard deviation
and mean of 5 and then these readings were averaged to give a dissimilarity matrix for both subjects
combined. In this matrix, 36 of the 45 elements had experimentally determined values. In all these
matrices unmeasured elements were filled with the mean of the remaining matrix elements. Owing to the
random target selection method, some matrix elements in the combined matrix had 4 separate trials,
whilst for 9 elements there was no experimental reading obtained.
Given the reduced number of trials, the MDS analysis is only really valid for the combined matrix
of both subject's data; the dissimilarity matrices for the separate subjects have 46% of the values missing
and under these conditions, given the noisy data, there is no likelihood of recovering structural
information. Nevertheless, for completeness, the 2 dimensional scaled solutions are given for the
separate subject data.
The three matrices of dissimilarity values, from each subject separately and from the two
combined, were permuted 25 times and the new matrices scaled. The stress values obtained from these
permutations were used to provide an estimate of the mean chance expectation for stress and a T-test was
applied to determine if the stress values for the experimental dissimilarity matrices were significantly
different from chance expectation.
15
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
STRESS VALUES FOR MDS MODELS OF THE PILOT STUDY.
Subject
Model
Dimension
Experiment
Stress
MCE
Stress
Probability
1 Tailed
AV
2
0.137
0.12
0.76
DK
2
0.137
0.132
0.58
BOTH
2
0.184
0.185
0.5
AV
3
0.055
0.052
0.6
DK
3
0.077
0.055
0.9
BOTH
3
0.096
0.1
0.38
These stress values do not differ significantly from chance. The values for the separate subjects are
higher than mean chance expectation, in the opposite direction to that hypothesized. This may be due to
the very high proportion of values in these matrices being mean values inserted to cover cases where no
experimental data was obtained. Given the small number of trials in relation to the 45 data points in the
dissimilarity matrix it is unsurprising that these chance level stress values occurred. The stress value,
0.14, of the comparison study described below is plausibly a minimum for the stress which might be
observed in the remote perception case. Noise added to the measured values of the dissimilarities would
increase the observed stress from this value towards the mean chance expectation value derived from
permuted dissimilarity matrices.
In spite of the high observed stress values in the pilot study, it is possible that the resulting MDS
structures will show organization which can be semantically interpreted. The 2 dimensional structures
for the combined subject data and for each subject separately are given below.
Approved For Release 2000/08/08 6 CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
?Dune coastline
Dune coastline
?
Grand Canyon
City
? ?Port
*City with canal
Poun to i n
.Waterfall
.Valley
'Lake
-2 -1 0 1 2
x
FIGURE 6. 2 DIMENSIONAL MDS STRUCTURE FROM BOTH SUBJECTS COMBINED.
City
Y 0
Mountain
?
Valley
?City with canal Fields
?Lake .Grand Canyon
?
Waterfall
x
FIGURE 7. 2 DIMENSIONAL MDS STRUCTURE FROM SUBJECT AV.
17
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
Grand Canyon
Y o
,city with canal 7
'Dune coastline
Port
Fields City ?
'Mountain
Valley Waterfall
? Lake
x
FIGURE 8. 2 DIMENSIONAL MDS STRUCTURE FROM SUBJECT DK.
FIGURE 9.3 DIMENSIONAL MDS STRUCTURE FROM BOTH SUBJECTS COMBINED.
Approved For Release 2000/08/081: C:IA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
s
As can be seen from Figure 6, the data from the subjects combined has resulted in a MDS
structure which appears to group the valley, waterfall and lake targets, all of which have prominent land-
water interfaces visible and are predominantly natural. However the dune coastline target also meets these
criteria and this is not in the same grouping. The urban scenes, the city, port and city with canal
targets, are also fairly close in the combined subject structure. We may therefore conclude that there is
weak evidence visible in this similarity data of discrimination of scenes containing land-water interfaces
from those without and between natural versus man-made scenes. The 3 dimensional structure for the
combined subject data given in Figure 9 also appears to group the urban scenes, but there is little
evidence of interpretable structure. These differentiations have not been tested statistically, but it appears
that more data of higher quality would be required to confirm them.
3 COMPARISON STUDY
In order to be able to compare the similarity values and semantic structure seen in the pilot study
with the corresponding data obtained by normal perception, a small study was performed using the same
target material, but unfortunately not the same subjects, as in the pilot study.
3.1 Method
Three subjects, all of whom had participated in other studies of remote perception but who
otherwise had no connection with this study participated. The subjects were requested to give a global
rating of similarity or difference on the scale shown in Figure 5 for all 45 pairs of the 10 targets used in
the pilot study. In a session of approximately one hour the subjects viewed the pairs of color pictures at
one meter distance and made a mark on the scale giving the perceived similarity. The target pairs were
presented in a random order.
3.2 Results
The dissimilarity ratings are given in Table 1 and the resulting 2 and 3 dimensional NMS
structures in Figures 1 and 2 respectively. The stress values were compared with those obtained by
permuting the experimental dissimilarity matrix randomly and computing an MDS structure exactly as
in the pilot study. Using 25 permutations to give chance stress estimates, the following statistics were
derived.
Approved For Release 2000/08/0 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
STRESS VALUES FOR MDS MODELS OF THE COMPARISON STUDY.
Model
Dimension
Experiment
Stress
MCE
Stress
Probability
1 Tailed
1
0.28
0.37
0.03
2
0.138
0.188
0.03
3
0.07
0.102
0.03
These stress results confirm that the similarity estimates produced in this experiment are
significantly structured. The resulting 2 and 3 dimensional MDS structures are interpretable in terms of
discrimination occurring due to man-made versus natural features and the presence or absence of land-water
interfaces and possibly pronounced vertical topography (see Section 1.3).
4 CONCLUSION
In this report we have described a novel method for investigating the semantics of remote
perception. The method appears to have merit in terms of its potential ability to discriminate aspects of
remote perception targets which may be preferentially perceived in this context. Therefore the method
may find use in the development of judging methods both by providing information on which descriptor
sets design can be optimized and also by judging remote perception efforts by similarity estimates
directly.
It is clear that the pilot study presented here has provided only very weak evidence of the
discrimination of targets by similarity estimates. However, given the large amount of quantitative data
required for the reliable determination of semantic structures and the unreliability of remote perception, it
might be expected that such a small study is unlikely to give clear results. Given a data set an order of
magnitude larger, useful results might well emerge.
Approved For Release 2000/08/08 CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
5.1 Appendix 1
The MDS algorithm used in the pilot study follows that due to Kruskal. 6,7 The program begins
by computing an initial configuration of points whose configuration is a linear function of the input data.
This is achieved by by a metric MDS method in which missing values in the input dissimilarity matrix
are replaced with the mean value of all elements of the matrix. Then the values are converted to distances
by adding a constant. A scalar products matrix B is then calculated by following the procedures in
Torgerson. 13 The initial configuration matrix for the non-metric MDS is then found from the
eigenvectors of B using the Young-Householder procedure. Nonmetric optimization then proceeds by
iterating the following sequence of steps: at the beginning of each iteration the configuration is
normalized to have zero centroid and unit dispersion. Next, Kruskal's DHAT (fitted) distances are
computed by a monotonic regression of distances onto data. The stress, S, is calculated and checked for
whether it has decreased sufficiently from the last iteration. If it has not, the negative gradient is
computed for each point by taking the partial derivatives of stress with respect to each dimension.
Points in the configuration are moved along their gradients by steps proportional to the derivatives and
the next iteration starts. After the last iteration the configuration is shifted so that its centroid lies at the
origin so that it has unit dispersion. For further details, see Wilkinson 18 and the cited references.
21
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
5.2 Appendix 2
DATA FROM PILOT STUDY. DISSIMILARITIES NORMALIZED TO UNITY STANDARD
DEVIATION AND MEAN OF 5.
Trial
Subject
Targets
Dissimilarity
Trial
Subject
Targets
Dissimilarity
32
AV
2 1
5.438
26
AV
7 6
5.541
40
DK
2 1
4.221
43
AV
7 6
4.458
9
AV
3 1
3.219
29
DK
8 2
5.182
2
DK
3 1
4.322
15
DK
8 3
5.991
23
AV
3 2
5.541
38
AV
8 5
4.509
56
DK
4 1
5.789
22
DK
8 5
4.272
39
DK
4 3
5.182
57
DK
8 6
6.193
13
DK
4 3
5.233
10
AV
8 7
6.108
30
DK
4 3
5.485
65
DK
8 7
3.210
51
DK
5 1
3.058
54
AV
9 1
3.684
8
AV
5 3
4.819
63
DK
9 2
4.727
14
DK
5 3
3.210
37
AV
9 3
3.116
5
DK
5 4
6.648
33
AV
9 3
5.593
41
AV
5 4
4.458
61
DK
9 3
6.244
27
AV
6 1
5.386
59
AV
9 4
3.735
47
AV
6 2
3.942
24
AV
9 5
6.108
7
DK
6 2
3.614
58
AV
9 5
4.303
52
DK
6 3
5.233
28
AV
9 7
3.735
1
DK
6 3
5.688
34
AV
9 7
5.438
50
DK
6 4
5.384
3
AV
10 1
5.696
21
DK
6 5
4.778
53
AV
10 1
5.696
45
DK
6 5
5.283
11
DK
10 1
3.159
19
DK
7 1
6.244
17
AV
10 2
6.418
42
AV
7 2
4.716
64
AV
10 2
6.057
62
AV
7 2
5.335
4
AV
10 5
5.335
36
AV
7 2
5.696
55
AV
10 6
3.632
12
AV
7 3
6.366
35
AV
10 7
4.406
20
AV
7 4
6.728
16
AV
10 7
6.366
18
DK
7 4
4.980
25
AV
10 7
3.735
31
AV
7 5
4.354
67
DK
10 7
4.778
46
DK
7 5
6.042
44
AV
10 8
5.335
66
DK
8 5
5.334
6
DK
10 8
4.778
60
DK
10 8
5.738
22
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
6 REFERENCES
1 . Daniel, T.C. and Ittelson, W.H., "Conditions for Environmental Perception Research: Comment
on "The Psychological Representation of Molar Physical Environments" by Ward and Russell,"
The Journal of Experimental Psychology: General, Vol. 110, No. 2, pp. 153 - 157 (1981).
2. Dunne B.J., JahnR.G. and Nelson R.D., "Precognitive Remote Perception," Technical Note
PEAR 83003, Princeton Engineering Anomalies Research Laboratory, Princeton University
School of Engineering/Applied Science, Princeton, New Jersey (1983).
3. Garling,T. "A Multidimensional Scaling and Semantic Differential Technique Study of the
Perception of Environmental Settings," Scandinavian Journal of Psychology, Vol. 17, pp. 323 -
332 (1976).
4. Garling, T. "The Structural Analysis of Environmental Perception and Cognition: A
Multidimensional Scaling Approach,"Environment and Behaviour, Vol. 8, No.3, pp. 385 - 390
(1976).
5. Jahn R.G. , Dunne B.J. and Jahn E.G., "Analytical Judging Procedure for Remote Perception
Experiments, "Journal of Parapsychology, Vol. 44, pp. 208 - 231 (1980).
6. Kruskal, J.B., "Multidimensional Scaling by Optimisation of Goodness of Fit to a Nonmetric
Hypothesis," Psychometrika, Vol. 33, pp. 469 - 506 (1964).
7. Kruskal, J.B., "Nonmetric Multidimensional Scaling: A Numerical Method," Psychometrika,
Vol. 29, pp. 115 - 129 (1964).
8. Lewis, T.G. and Payne, W.H., "Generalized Feedback Shift Register Pseudorandom Number
Generator," Journal of the Association for Computing Machinery, Vol. 20, No. 3, pp. 456 - 463
(1973).
9. May, E.C., "Intuitive Data Sorting: An Informational Model of Psychoenergetic Functioning,"
SRI International , Menlo Park, California (1986).
10. Mehrabian, A. and Russell, J.A., "A Verbal Measure of Information Rate for Studies in
Environmental Psychology," Environment and Behaviour, Vol. 6, No.2, pp. 231 - 251 (1974).
11. Puthoff H.E. and Targ R., "A Perceptual Channel for Information Transfer over Kilometer
Distances: Historical Perspective and Recent Research," Proceedings IEEE, Vol. 54, pp. 329 -
354 (1976).
12. Spence, I. & Ogilvie, J.C., "A Table of Expected Stress Values for Random Rankings in
Nonmetric Multidimensional Scaling," Multivariate Behavioural Research, Vol. 8, pp. 511 -
517 (1973).
13. Torgerson, W.S., Theory and Methods of Scaling, John Wiley & Sons, New York (1958).
14. Tversky, A., "Features of Similarity," Psychological Review, Vol. 84, pp. 327 - 352 (1977).
15. Tversky, A. and Gati, I., "Similarity, Separability and the Triangle Inequality,"
Cognition and Categorization, Vol. 1978 (1978).
Approved For Release 2000/08/&3: CIA-RDP96-00789R002200130001-3
 Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3
16. Ward, L.M. and Russell, J.A., "The Psychological Representation of Molar Physical
Environments," The Journal of Experimental Psychology: General," Vol. 110, No. 2, pp. 121 -
152 (1981).
17. Ward, L.M., "Multidimensional Scaling of the Molar Physical Environment," Multivariate
Behavioural Research, Vol. 12, pp. 23 - 42 (1977).
18. Wilkinson, Leland., SYSTAT: The System for Statistics., Evanston, IL: SYSTAT Inc. (1986).
19. Young, F.W., "Nonmetric Multidimensional Scaling: Recovery of Metric Information,"
Psychometrika, Vol. 35, No. 4, pp. 455 - 473 (1970).
24
Approved For Release 2000/08/08 : CIA-RDP96-00789R002200130001-3