Social Cognition in Silica
A ‘Theory of Mind’ for Socially Aware Artificial Minds
Michael Harr
´
e
Complex Systems Research Group, Faculty of Engineering and IT, The University of Sydney, Sydney, 2006, NSW, Australia
Keywords:
Theory of Mind, Artificial Intelligence, Social Networks, Game Theory.
Abstract:
Each of us has an incredibly large repertoire of behaviours from which to select from at any given time,
and as our behavioural complexity grows so too does the possibility that we will misunderstand each other’s
actions. However, we have evolved a cognitive mechanism that allows us to understand another person’s
psychological space: their motivations, constraints, plans, goals and emotional state and it is called our ‘Theory
of Mind’. This capability allows us to understand the choices another person might make on the basis that the
other person has their own ‘internal world’ that influences their choices in the same way as our own internal
world influences our choices. Arguably, this is one of the most significant cognitive developments in human
evolutionary history, along with our ability for long term adaptation to familiar situations and our ability to
reason dynamically in completely novel situations. So the question arises: Can we implement the rudimentary
foundations of a human-like Theory of Mind in an artificial mind such that it can dynamically adapt to the
likely decisions of another mind (artificial or biological) by holding an internal representation of that other
mind? This article argues that this is possible and that we already have much of the necessary theoretical
foundations in order to begin the development process.
1 INTRODUCTION
One of the key drivers of work on the development
of artificial human-like reasoning has focused on how
we come to understand the inanimate world. For ex-
ample, it is not uncommon to discuss cognitive devel-
opment almost solely in terms of the cognitive struc-
tures relating to inanimate objects (see (Tenenbaum
et al., 2011) for an example, where work on Theory
of Mind is mentioned in Open Questions). While the
learning of inanimate relationships is an important di-
rection to explore, it is only one piece of the puzzle
of cognition. In contrast, the focus of this work is
on our ability to understand interpersonal animate re-
lationships, an ability that appears to be equally old
and important in evolutionary terms as other forms of
comprehending the world.
Earlier work in anthropology has revealed some
striking facts about our neuro-cognitive architecture
and the role it plays in our social development as a
species. In 1992 Robin Dunbar put forward the hy-
pothesis that our neocortex grew in size in order to ac-
commodate the cognitive pressures imposed upon us
by our growing social group size. One of the notable
predictions to come out of this work is that humans
should have a social group size of approximately 150
individuals (Dunbar, 1992), this has been called Dun-
bar’s Number. The relationship between the neo-
cortex size and social group size is called the So-
cial Brain Hypothesis (Dunbar and Shultz, 2007) and
over the last 20 years considerable evidence across
many primate and non-primate species has been col-
lected supporting the hypothesis (Kudo and Dunbar,
2001; Shultz and Dunbar, 2007; Lehmann and Dun-
bar, 2009). Perhaps most significantly, recent studies
have shown that this hypothesis is true at the individ-
ual as well as at the species level. In a study pub-
lished last year (Powell et al., 2012) by Dunbar and
colleagues it was shown that in individual humans,
the size of their social network was linearly related to
the neural volume of the orbital prefrontal cortex. A
secondary but critical finding of the same study is that
neural capacity is necessary but not sufficient for large
social networks, the subjects also needed to have de-
veloped the psychological skills necessary for under-
standing another persons point of view. This cogni-
tive skill, called Theory of Mind (ToM (Baron-Cohen
et al., 2000)), enables us to understand that other peo-
ple have mental states and that these states might be
different from our own. This enables us to better
manage our social relationships and maintain a larger,
more complex and more diverse set of social relation-
657
Harré M..
Social Cognition in Silica - A ‘Theory of Mind’ for Socially Aware Artificial Minds.
DOI: 10.5220/0004917606570662
In Proceedings of the 6th International Conference on Agents and Artificial Intelligence (ICAART-2014), pages 657-662
ISBN: 978-989-758-015-4
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
ships than any of our primate relatives.
Following on from these studies, the aim of this
article is to introduce a simple model of social inter-
actions between artificial agents that imply a view of
the agent as filtered through the strategic perspective
of another agent, this is called a strategic Theory of
Mind (Harr
´
e, 2013). This is then extended to interac-
tions of multiple agents in a social network such that a
single agent’s decision-making process is influenced
by those in closest proximity in their social network,
but these closest relationships are in turn influenced
by second order relationships that are not directly re-
lated to the first agent. The result is an extension of
strategic ToM to a socio-strategic ToM and social net-
works in general.
2 SPECIFIC FUNCTIONAL
ROLES OF THEORY OF MIND
Theory of Mind research looks at how we are able
to reason based on an internal representation of how
an individual believes other people’s minds operate in
general and then to use this representation to under-
stand how specific contexts influence another individ-
ual’s actions. In strategic interactions, such as eco-
nomic game theory, understanding another person’s
state of mind has the direct and obvious advantage
of benefiting in terms of increased payoffs (Bhatt and
Camerer, 2005), but these benefits extend to every as-
pect of our lives, to how we teach children, collab-
orate in scientific research, empathise with the less
fortunate and how the economic division of labour al-
lows us to divide tasks according to the specific skills
and abilities of each individual.
In order to understand the neural foundations of
our ToM, recent progress has been made in the neuro-
imaging of human subjects carrying out ToM related
tasks. A complex network of brain regions have
been revealed that are activated during any cogni-
tive task that involves thinking of another person’s
state of mind or even social interactions with animate
rather than inanimate agents. Focusing specifically
on understanding and internally representing the men-
tal states of others, two of the most important func-
tional properties of these brain networks are the abil-
ity to recognise that people, unlike other things in the
world, have mental states that include thoughts, feel-
ings, constraints, goals and perceptions and the de-
velopment of an internal model of how these men-
tal states influence the decisions they make within a
specific environmental context (Lieberman, 2007). In
this article, the focus will be on person A thinking of
the environmental context in which person B is mak-
ing decisions, and B’s environment will be a social en-
vironment (that may include A). Note that this is only
a subset of the possible contexts in which B could be
making a decision and A might still find it useful to
have a ToM for B in such contexts, but this is not the
focus of this article.
A special case worth highlighting is perspective
taking, which most commonly refers to understand-
ing the sensory perception of another person, for ex-
ample that another person sees something different
to what you can see. Humans can solve perceptual
perspective-taking tasks using visuo-spatial reasoning
without the need of a ToM mechanism (Zacks and
Michelon, 2005). However there is a broader mean-
ing to perspective taking that includes adopting or
considering another person’s psychological perspec-
tive, and this is sometimes understood as being syn-
onymous with empathy (Lieberman, 2007) (and so
sometimes is called cognitive empathy (Lamm et al.,
2007)). From this point of view adopting another per-
son’s perspective is equivalent to a person trying to
place themselves in the same psychological space as
the other person, including emotions, constraints etc.
and this is called the simulation theory of ToM (Gold-
man, 2005) (contrast with theory-theory ToM). This
article proposes a simplified form of the simulation
theory of ToM: an artificial mind can potentially con-
tain a model of another agent’s psychological per-
spective (either artificial or human), and can use this
perspective to improve their decision-making in so-
cial contexts.
2.1 A ‘Game Theory of Mind’
Arguably one of the most significant insights to come
from economics is to ask the central question: How
do people make decisions in the context of other peo-
ple’s decisions? This strikes close to issues central
to our ToM, if one person understands that another
person’s actions will change the reward they will re-
ceive, then understanding the way in which that other
person chooses their actions would be an invaluable
tool. From this point of view, without comprehend-
ing another person’s inner cognitive workings when
the value of a reward depends upon the other’s de-
cisions a vast world of cooperative and competitive
advantage is lost to us. Such reasoning requires in-
dividuals to account for how other’s view each of the
likely decision’s everyone else will make, and in do-
ing so they collectively change the decision-making
patterns of the collective.
In a similar vein, decision-making has been mod-
elled across large populations using stochastic differ-
ential equations in order to explain their choices in
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
658
economic games. In a notable study, at an economic
conference on game theory, the participants were
asked to play a game called the Traveler’s Dilemma
for real monetary rewards (Goeree and Holt, 1999;
Anderson et al., 2004). While the participants did
not play the strict equilibrium strategies predicted by
classical economics, they did make decisions that col-
lectively were in agreement with a form of bounded
rationality equilibrium described by statistical evo-
lutionary equations and whose stationary states are
again reminiscent of some probabilistic models used
in Artificial Intelligence (AI) and theoretical psychol-
ogy.
In the final study considered here, Yoshida et
al. (Yoshida et al., 2008) recently proposed what they
called a Game Theory of Mind that uses bounded re-
cursion (of the sort: I’m thinking of you thinking of
me thinking of you etc. truncated at a certain level)
and value functions attributable to different players in
order to model depth of strategic reasoning.
2.2 A Stochastic Model of
Decision-making
Conventional game theory begins with a number of
players and the rewards they receive for their joint ac-
tions, in the simplest case there are two players (here
called A and B), each of which has two choices and
the payoff matrices are given by:
α =
α
1
1
α
1
2
α
2
1
α
2
2
, (1)
β =
β
1
1
β
1
2
β
2
1
β
2
2
. (2)
The expected utility to each player is given in terms of
the joint probability p(α
i
)q(β
j
) = p(A
i
)q(B
j
) = p
i
q
j
.
Here, A denotes the player, A
i
denotes a decision vari-
able that can take two different value, either α
1
or
α
2
denoting the first or second row in equation 1,
and the A
i
are sometimes called cumulative decision
variables (see below for where this comes from) and
p(A
1
) = A
1
/(A
1
+ A
2
) etc. (equiv. for B), so:
E
a
(u) =
i, j
p
i
q
j
α
i
j
, E
b
(u) =
i, j
p
i
q
j
β
j
i
(3)
A further set of equations are needed as well, called
the conditional expected utilities (Wolpert et al.,
2012):
E
a
(u|α
i
) =
j
q
j
α
i
j
, E
b
(u|β
j
) =
i
p
i
β
j
i
(4)
in which the expected utility to α is conditional upon
α fixing their choice to α
i
(similarly for β). Us-
ing these expressions we can represent how a single
player (A in what follows) models the theirs and their
opponents incremental changes in the decision vari-
ables during time interval dt in terms of a determinis-
tic drift term and a stochastic diffusion term (Bogacz
et al., 2006):
dA
i
= (α
i
1
e
B
1
+ α
i
2
e
B
2
)dt + σ
i
A
dW
A
(5)
d
e
B
j
= (
e
β
i
1
A
1
+
e
β
i
2
A
2
)dt + σ
j
B
dW
B
(6)
Figure 1: The neuro-cognitive development of a ToM by
player A. In each panel A
1
signals before A
2
and A
1
then
laterally inhibits A
2
from signalling (inhibition connections
not shown). A: Two cell assemblies encode stochastic deci-
sion variables A
1
and A
2
that take early signals from other
regions of the brain. B: Two new cell assemblies are formed
that encode the decision variables B
1
and B
2
of another
player, these decision variables influence the A
i
decision
variables via black dashed connections. C: The A
i
are now
reciprocally connected to the B
j
through red dashed con-
nections such that dynamic changes in the activity of the
A
i
cell assemblies during the decision making process are
reflected in the activity of the B
j
cell assemblies.
Note the tilde in these equations indicating that it
is the A player’s estimate of the B player’s utility and
SocialCognitioninSilica-A'TheoryofMind'forSociallyAwareArtificialMinds
659
decision variable B, in game theory the utility is com-
mon knowledge between players, but in representing
the other player’s decision process each player has to
estimate the utility the other player will gain from the
interaction. Equations 5 and 6 represent a model of
how A has neurally encoded the constraints, their esti-
mates of Bs incentives as well as their decision mak-
ing process (it is the same as As). In this representa-
tion of stochastic neural signal accumulation, the dA
i
are only implicitly connected to each other through
their connection with the B
j
(no A
i
term appears on
the right hand side of equation 5), however each A
i
interacts with both B
j
terms. This is how the neural
network of A looks once it has fully developed, but the
intermediate steps to this final form can be described
as:
A. dA
i
= µ
i
dt + σ
i
A
dW
A
(7)
B. dA
i
= (α
i
1
e
B
1
+ α
i
2
e
B
2
)dt + σ
i
A
dW
A
(8)
d
e
B
j
=
e
µ
j
dt + σ
j
B
dW
B
(9)
where the letter labels on the left reflect the panel la-
bels of Figure 1 and panel C. is modelled by equa-
tions 5 and 6. The µ
i
term of equation 7 represents
the weight the player attributes to each of their two
options, because the player is not accounting for the
other player’s strategies at this point, a plausible (but
by no means the only) strategy is to take the average
of the two payoffs available for each choice for ex-
ample: µ
i
= (α
i
1
+ α
i
2
)/2 (similarly for Bs µ
j
). In
terms of strategic thinking (see as an example (Cori-
celli and Nagel, 2009)), option A. is level 0 thinking,
no account is made of the other player’s choices, the
choice of µ
i
= (α
i
1
+ α
i
2
)/2 implies A assumes the
other player equally weights their choices, but there
are obviously many other alternative formulations of
µ
i
. Option B. is level 1 thinking, some account is
made of Bs strategy, but no attempt is made by A
to adjust their interpretation of Bs strategy by ac-
counting for how B might be strategically weighting
their choices based upon As likely strategy. Finally,
equations 5 and 6 represent A accounting for their es-
timation of the weighted strategies of B where Bs
weighted strategies accounts for As weighted strate-
gies (level 2 thinking).
It is not easy to see that there is necessarily a solu-
tion to these equations such that an equilibrium in the
dynamics might be achieved. However, it has recently
previously been shown that providing the drift terms
are linear in the decision variables (the terms that pre-
cede dt in equations 5- 6) there is a guaranteed set of
equilibrium probabilities given by:
p(A
i
) e
γ
a
E
a
(u|α
i
)
(10)
p(B
j
) e
γ
b
E
b
(u|β
j
)
(11)
where the γ
a
and γ
b
terms are proportional to the noise
terms σ in equations 5 and 6 and the remainder of
the terms in the exponents are simply the conditional
expected utilities of equation 4. The interpretation of
this form is that if A holds strategy A
i
fixed then the
probability of choosing strategy A
i
is proportional to
the exponentiation of the utility condition on A
i
being
fixed based upon As estimate of the distribution over
strategies B will choose.
In order to see how equations 10 and 11 can be
thought of as social perspective taking, simplify these
probabilities to only one variable for each player:
Q
a
= 1 2p(A
1
) [1, 1] and Q
b
= 1 2q(B
1
)
[1,1]. Now rewrite equations 10 and 11 in an ex-
plicit form in terms of an equilibrium involving only
Q
a
for A:
Q
a
= 1
2e
γ
a
E
a
(u|α
1
)
e
γ
a
E
a
(u|α
1
)
+ e
γ
a
E
a
(u|α
2
)
(12)
= tanh(
γ
a
2
(E
a
(u|α
2
) E
a
(u|α
1
))) (13)
= tanh(
γ
a
2
(z
a
0
+ z
a
1
Q
b
)) (14)
= tanh(
γ
a
2
(z
a
0
+
z
a
1
(tanh(
γ
b
2
(z
b
0
+ z
b
1
Q
a
))))) (15)
The z terms are reduced constants derived from pay-
off matrices 1 and 2. Note that Q
a
can be written
as a function of Q
b
: Q
a
= F
a
(Q
b
) where F
a
(·) is
As decision model with As γ and z parameters, cf.
equation 14 and likewise Q
b
= F
b
(Q
a
). However,
these are implicit self-consistent equations for each
player’s choices: Q
a
= F
a
(F
b
(Q
a
)) (cf. equation 15)
and Q
b
= F
b
(F
a
(Q
b
)).
2.3 Social Networks, Decisions and ToM
Between players A and B a minimal 2-person social
network has been described, one of mutual and self
consistent comprehension between the two players
(when in equilibrium) based upon an accurate men-
tal model each has of the other. This small social
network can be expanded to multiple agents, to do
so label each new agent C, D, E etc. and like A
and B these new comers only have two options in a
game theory-like setting to choose from so that Q
c
,
Q
d
, Q
e
etc. can be defined similarly to Q
a
= F
a
(Q
b
).
Taking the perspective of how A needs to internally
model their social network in order to make decisions
that are ‘in equilibrium’, As model will depend upon
the local topology of their social network. For exam-
ple if A is connected to four other people and these
people do not know each other or anyone else other
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
660
than A, then A can represent them as independent:
Q
a
= F
a
(F
b
(Q
a
),F
c
(Q
a
),F
d
(Q
a
),F
e
(Q
a
)), i.e.
Q
a
= tanh
h
γ
a
2
(z
a
0
+
z
a
1
(tanh(
γ
b
2
(z
b
0
+ z
b
1
Q
a
))) +
z
a
2
(tanh(
γ
c
2
(z
c
0
+ z
c
1
Q
a
))) +
z
a
3
(tanh(
γ
d
2
(z
d
0
+ z
d
1
Q
a
))) +
z
a
4
(tanh(
γ
e
2
(z
e
0
+ z
e
1
Q
a
))))
i
(16)
where the z and γ notation has been extended to these
new players, see Figure 2, B. Alternatively, given
three agents that all know each other and interact with
each other, As model of their local social network
would be Q
a
= F
a
(F
b
(F
c
(Q
a
))), i.e.
Q
a
= tanh
h
γ
a
2
(z
a
0
+
z
a
1
(tanh(
γ
b
2
(z
b
0
+
z
b
1
(tanh(
γ
c
2
(z
c
0
+
z
c
2
(tanh(
γ
d
2
(z
d
0
+ z
d
1
Q
a
)))
i
(17)
see Figure 2, C. In this example it would not be a com-
plete representation of the local network topology for
A to simply have accounted for B and C as though
these two people were independent of each other, in
such a model: Q
a
= F
a
(F
b
(Q
a
),F
c
(Q
a
)), but this does
not accurately reflect how A needs to consider the way
in which B and C adjust their decisions based upon the
connection they have with each other, as this connec-
tion indirectly influences how A needs to consider the
choices they make. More generally, in the case of B.
of Figure 2, while A is influenced by B (and C, D and
E), B might themselves be connected to other players
in their local network and these players are only indi-
rectly related to A through the influence they have on
Bs decisions.
3 CONCLUSIONS
How do agents, artificial or biological, coordi-
nate their actions in such a way that their collec-
tive decision-making is better than their individual
decision-making? People are both good and bad at
such aggregation: science is based on individuals co-
operating and competing in order to produce a body of
knowledge far greater than any one individual could
achieve, but the coordinated behaviour of traders in
Figure 2: The internal representations A has of the
other agents in different network topologies. Each
topology has a distinctive representation when As
choices are in equilibrium: A: Q
a
= F
a
(F
b
(Q
a
)),
B: Q
a
= F
a
(F
b
(Q
a
),F
c
(Q
a
),F
d
(Q
a
),F
e
(Q
a
)), C: Q
a
=
F
a
(F
b
(F
c
(Q
a
))).
financial markets can lead to billions of dollars be-
ing lost in a market crash. More importantly for ar-
tificial agents, what are the theoretical foundations of
cooperation and competition that can lead to simple
agents making better decisions collectively than could
be achieved through every individual acting indepen-
dently of one another? So how do we engineer true
collective intelligence in societies of artificial agents?
One approach is outlined in this work: individual
agents make their decisions based on what they can
each individually and objectively know (or can find
out) about the environment as well as how they be-
lieve other agents will make their choices and how
these choices subsequently impact on the quality of
their own decisions. In people, such cooperative di-
vision of labour leads to specialisation and expertise
amongst a population that is able to solve problems
that no individual could solve alone, replicating this
in a collection of artificial intelligences will open up
the possibilities of a symbiosis of bio-silica commu-
nities, where artificial agents are dynamically respon-
sive to the internal mental states of people, enhancing
the quality of our decision-making.
SocialCognitioninSilica-A'TheoryofMind'forSociallyAwareArtificialMinds
661
REFERENCES
Anderson, S. P., Goeree, J. K., and Holt, C. A. (2004).
Noisy directional learning and the logit equilibrium.
The Scandinavian Journal of Economics, 106(3):581–
602.
Baron-Cohen, S. E., Tager-Flusberg, H. E., and Cohen,
D. J. (2000). Understanding other minds: Perspec-
tives from developmental cognitive neuroscience . Ox-
ford University Press.
Bhatt, M. and Camerer, C. F. (2005). Self-referential think-
ing and equilibrium as states of mind in games: fmri
evidence. Games and Economic Behavior, 52(2):424–
459.
Bogacz, R., Brown, E., Moehlis, J., Holmes, P., and Cohen,
J. D. (2006). The physics of optimal decision making:
a formal analysis of models of performance in two-
alternative forced-choice tasks. Psychological review,
113(4):700.
Coricelli, G. and Nagel, R. (2009). Neural correlates of
depth of strategic reasoning in medial prefrontal cor-
tex. Proceedings of the National Academy of Sciences,
106(23):9163–9168.
Dunbar, R. (1992). Neocortex size as a constraint on
group size in primates. Journal of Human Evolution,
22(6):469–493.
Dunbar, R. I. and Shultz, S. (2007). Evolution in the social
brain. science, 317(5843):1344–1347.
Goeree, J. K. and Holt, C. A. (1999). Stochastic game
theory: For playing games, not just for doing the-
ory. Proceedings of the National Academy of Sciences,
96(19):10564–10567.
Goldman, A. I. (2005). Imitation, mind reading, and sim-
ulation. Perspectives on Imitation: Imitation, human
development, and culture, 2:79.
Harr
´
e, M. (2013). The neural circuitry of expertise: Percep-
tual learning and social cognition. Frontiers in Human
Neuroscience, 7:852.
Kudo, H. and Dunbar, R. (2001). Neocortex size and so-
cial network size in primates. Animal Behaviour,
62(4):711–722.
Lamm, C., Batson, C. D., and Decety, J. (2007). The neural
substrate of human empathy: Effects of perspective-
taking and cognitive appraisal. Journal of cognitive
neuroscience, 19(1):42–58.
Lehmann, J. and Dunbar, R. (2009). Network cohesion,
group size and neocortex size in female-bonded old
world primates. Proceedings of the Royal Society B:
Biological Sciences, 276(1677):4417–4422.
Lieberman, M. D. (2007). Social cognitive neuroscience: a
review of core processes. Annual Review of Psychol-
ogy, 58:259–289.
Powell, J., Lewis, P. A., Roberts, N., Garc
´
ıa-Fi
˜
nana, M.,
and Dunbar, R. (2012). Orbital prefrontal cortex
volume predicts social network size: An imaging
study of individual differences in humans. Proceed-
ings of the Royal Society B: Biological Sciences,
279(1736):2157–2162.
Shultz, S. and Dunbar, R. (2007). The evolution of the so-
cial brain: Anthropoid primates contrast with other
vertebrates. Proceedings of the Royal Society B: Bio-
logical Sciences, 274(1624):2429–2436.
Tenenbaum, J. B., Kemp, C., Griffiths, T. L., and Goodman,
N. D. (2011). How to grow a mind: Statistics, struc-
ture, and abstraction. science, 331(6022):1279–1285.
Wolpert, D. H., Harr
´
e, M., Olbrich, E., Bertschinger, N.,
and Jost, J. (2012). Hysteresis effects of changing the
parameters of noncooperative games. Physical Review
E, 85(3):036102.
Yoshida, W., Dolan, R. J., and Friston, K. J. (2008).
Game theory of mind. PLoS computational biology,
4(12):e1000254.
Zacks, J. M. and Michelon, P. (2005). Transformations of
visuospatial images. Behavioral and Cognitive Neu-
roscience Reviews, 4(2):96–118.
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
662