Download Wayward Agents in a Commuting Scenario

Wayward Agents in a Commuting Scenario
(Personalities in the Minority Game)
Ana L.C. Bazzan
Rafael H. Bordini
Gustavo Kuhn Andrioti
Rosa M. Vicari
Institute for Informatics
Universidade Federal do Rio Grande do Sul (UFRGS)
Porto Alegre, Brazil
fbazzan, bordini, fgka, [email protected]
Joachim Wahle
Physics of Transport and Traffic
Gerhard-Mercator-University, Duisburg, Germany
[email protected]
Abstract
particularly in economics. It has also been the subject of
recent studies in statistical physics, e.g., [8, 13].
In our daily lives we often have to face binary decisions
where we seek to take the minority’s choice, e.g., in traffic
scenarios where we have to choose between similar alternative routes. We have studied agent coordination mechanisms in a binary decision model introduced in the literature recently: the Minority Game. Extending this model, we
have introduced personalities which model certain types of
human behaviour. Different populations of these personalities have been simulated. It was found that there is one
personality which performs better than the average: the
wayward personality. We have also extended the concept
of Hamming distance (a standard probabilistic measure for
the difference between strategies) to account for our personalities; this notion is useful for qualitatively explaining
or predicting results. Our work gives insights into the impact of commuters’ behaviours and it addresses relegated
issues in traditional traffic simulation.
The idea of rewarding the decision that is made by the
minority of the players (or agents) is interesting in many
scenarios. In our work on using multi-agent simulations of
traffic, minority game is clearly useful: choosing routes that
are least busy is thoroughly rewarding. This is the main motivation for the work reported in this paper. In effect, we can
rephrase the pub scenario above in terms of choosing a certain route for commuting only to find out that it is awfully
jammed, and then thinking to oneself: “I wish I had taken
the alternative route”.
1. Introduction
Who has not been through the situation where one has
gone to one’s favourite pub only to find out that it happens
to be overcrowded that night, and then regretting not to have
stayed home? That is precisely the motivation for the recent
minority game (MG) — a game where the decision that is
made by the minority of the players is rewarded. It was
first referred to as the El Farol Bar Problem (EFBP) [1], and
this game is presently booming in many application areas,
We have developed a set of simulations of the EFBP
where we have considered agents with personalities (as in
[7]), which are meaningful combinations of what is usually
called strategies in the game. These personalities are related
to commuters’ behaviours. We aim at drawing conclusions
as to what types of behaviour are more rewarding to commuters and beneficial to the traffic as a whole. In fact, our
results apply quite generally in the area of Multi-Agent Systems. Agents typically compete for resources available in
their environment: to avoid sought-after resources will be
clearly advantageous to any agent.
Our work has concentrated on anthropomorphic aspects
in traffic simulation, which microeconomics-based simulations usually relegate. In [4], we have produced simulations
of the Iterated Prisoner’s Dilemma (IPD) where we have introduced the idea of groups of agents, and tested ideas from
Ridley’s [15] theories of moral sentiments. In [5] we have
discussed the necessity of changing the paradigm of modelling drivers’ behaviour in traffic flow systems. It turns out
that information can have a serious impact on stability of
traffic conditions [16].
The next section reviews the fast growing literature on
the EFBP. Section 3 presents further the motivations for
this work using the EFBP applied to a traffic scenario. We
then present our simulations of the EFBP where some ideas
related to personalities have been considered in Section 4.
The results are analysed in Section 5. Finally, Section 6 discusses several aspects of this work and its future directions,
and provide concluding remarks.
2. Literature Review
Microeconomics and game-theory assume human behaviour to be rational and deductive – deriving a conclusion
by perfect logical processes from well-defined premises.
But this assumption does not hold especially in interactive
situations like the coordination of many agents. There is no
a priori best strategy since the outcome of a game depends
on the other players. Therefore, bounded and inductive rationality (i.e. making a decision based on past experience)
is supposed to be a more realistic description.
In this spirit Arthur introduced in 1994 a coordination
game [1] called the El Farol Bar Problem (EFBP). (The
name is inspired by a bar in Santa Fe.) Every week n players wish to visit the bar El Farol attracted by live music. Up
to a certain threshold of customers T the bar is very comfortable and enjoyable. If it is too crowed, it is better to stay
home. The payoff of the game is clear: if the number of
visitors is less than the threshold T these are rewarded, otherwise those who stayed home are better off. In the original
work, they used n = 101 and T = 60, but arbitrary values
of n and T have also been studied [13, 12].
The players can only make predictions about the attendance for the next time based on the results of the previous m weeks; this is the basis for strategies for playing the
game. Now, there is a huge set of possible strategies and
every player possesses k of them. For the decision whether
to go or to stay home, the player always selects the strategy
which predicts the outcome of the last weeks most accurately. After every week the player learns the new attendance and thus evaluates his strategies. By means of computer simulations, it was realised that the mean attendance
always converges to the threshold T = 60 [1].
Later on the EFBP was generalised to a binary game by
Challet and Zhang [8], the so-called Minority Game (MG).
An odd number n of players has to choose between two
alternatives (e.g., yes or no, or simply 0 or 1). With a memm
ory size m there are 22 possible strategies. Each player
has a set S of them which are chosen randomly out of the
whole set. The strategies are chosen as before and the minority group wins. In the simplest version of the game, these
players are rewarded one point. Since the step function is
a very simple payoff, other functions which favour for instance, smaller minorities were studied by several authors
[8, 9, 13].
The MG and the EFBP are gradually becoming a
paradigm for complex systems and have been recently studied in detail1 . We will refer briefly to some of the basic
results. In their original work, Challet and Zhang have systematically studied the influence of the memory size m and
number of strategies S on the performance of the game.
They concluded that the mean attendance always converges
to n=2 but for larger m there are less fluctuations, which
means the game is more efficient.
Additionally, different temporal evolution processes
were studied. It has been shown that such evolutionary processes lead to stable, self-organised states [8]. Some analytical studies showed that there is a phase transition to an
efficient game if the parameter = 2m =n is approximately
one. This is strongly related to the value of the Hamming
distance [9, 10]. The Hamming distance is the probability
that the decision of two strategies differ from each other, a
valuable concept used later on in this paper.
Finally, as we write up the final version of this paper,
Edmonds has just published on the subject of the EFBP
[11]; he investigates the emergence of heterogeneity among
agents in a simulation. His paper tackles evolutionary learning as well as communication among agents, which might
lead to a differentiation of roles at the end of the run. In
a sense, this is also what we aim in the present paper, except that we have based the design of our agents to represent
known strategies or class of agents which may be found in
real societies.
Another characteristic of Edmonds’s paper which resembles ours is his focussing on the description of actual human
behaviour. This compromise is clear when he describes the
structure of their agents (ability to communicate and learn).
It seems that, due to this quite heavy approach, he has to
deal with a very reduced number of agents (ten), which
poses the question of whether the influence of a single agent
in the system is not excessive. Also, the fact that he has
based his conclusions on 11 to 20 runs only gives a hint of
how hard the problem of simulating interactions of this kind
is.
Our approach is to be located somewhere between the
heavily cognitive-focussed approach by Edmonds and the
one by Challet and Zhang, which uses statistical physics
and so does not account for human motivations.
3. Motivation: A Commuting Scenario
The EFBP is just a metaphor for binary decisions which
happen in our daily lives very often [12]. In traffic flow such
1 For a collection of papers and pre-prints on the Minority Game, see
the web page at URL http://www.unifr.ch/econphysics/.
situations are of great interest because road users are often
confronted with binary route choice problems. In particular,
decisions where a set of agents compete daily are important
since they allow the investigation of evolutionary patterns.
To discuss this problem in more detail let us assume the
following situation. A set of commuters want to go from
home to work. These two locations are connected by two
routes, A and B, which have nearly the same capacity and
link travel time2 . Everyday the commuters have to decide
which route to select.
Another problem for the commuters is that they do not
know whether their choice was good or not because they do
not have information about the travel time on the alternative route. They have to decide on the basis of their past
experience. From daily experience, one can draw the conclusion that road users somehow reach a stable equilibrium,
i.e., traffic conditions on the roads are nearly the same everyday3. In particular situations, this stability can vanish
due to traffic messages, as information adds a new degree
of freedom to this simple game. That is so because people
have different ways to react to suggestions: it depends on
their personalities [16].
When choosing which particular route to take, there are
many aspects which can make a commuter’s life easier, e.g.,
choosing the shortest route, the one where the quality of the
road is better, the one that has less traffic lights, and even
subjective things like the one with a pleasant landscape. But
other things being equal, selecting the route that fewer people have chosen, clearly makes life a lot easier. In other
words, it is a complex problem to assure that road users are
well distributed along similar routes.
It is precisely in such complex systems where this work
can be useful. It is clearly a difficult problem to make decisions in such dynamic scenarios, as there is usually a huge
number of commuters making decisions on the fly, based on
so many variables, and where the slightest detail (e.g., radio broadcast of traffic conditions) can have unpredictable
consequences. The scenario is very tempting for traditional
decision-theoretic approaches, but these notoriously neglect
the more subjective aspects of commuters’ thought processes. More than that, they fail to consider intra-cultural
variation, to put it in anthropological terms [6]. By that we
mean that decisions are profoundly dependent on personalities, and thus are idiosyncratic in any considered community of commuters.
In order to account for commuters idiosyncratic decisions, we add to our simulations the idea that agents
have personalities, which are sets of meaningful (so-called)
2 Note that equal capacity is only assumed for simplicity. The routes
can also have different capacities. This is analogue to the use of thresholds
in the EFBP and the MG.
3 As we have observed in a case study for downtown Duisburg
http://www.traffic.uni-duisburg.de/OLSIM.
strategies. An agent of a certain personality selects randomly among the set of strategies building up its personality; this idea is expounded later in Section 4. With this
work, we aim at investigating what is the impact of certain
personalities in such traffic scenarios as well as the performance that particular personalities have in such scenarios.
This work can also bring us to a better understanding of the
dynamics of traffic and in future work come to provide directives for traffic organisation that can be of use to those in
charge of transport policies.
In effect, our work applies generally to multi-agent coordination: we find that the complexity inherent in traffic systems is a good case study for the general problem
of coordinating multiple self-interested agents, a research
line already tackled by us in [2, 3]. Therefore, results that
are meaningful in such traffic environments can certainly be
extrapolated to general coordination mechanisms in multiagent systems.
4. Simulation of the Personalities
Similarly to the standard approach followed in the Minority Game (MG), n players or agents choose a side to be
in. According to the collective choice, a reward (1 point) is
paid to all agents who happen to belong to the minority, i.e.,
to those who have chosen the side having less agents. The
parameter m (memory) introduced by Challet and Zhang
(1997) determines the length of the “history” a player considers in order to make the actual choice. An array of size
m keeps the history of which side won the last m games.
This way, m also determines the (finite) number of tactics
an agent may choose. In fact, the term “strategy” is used in
the literature. However, we prefer to call this tactics rather
than strategies, for a strategy is much more a plan or a policy, i.e., a previous level in the decision-making process.
Random choice is far from being what people actually do.
Rather, they have strategies and make a choice of tactics
according to them. On their turn, such strategies are determined by their personalities.
Besides, we find it implausible that ordinary people –
drivers in our scenario – may deal with a set of more than
a dozen of possibilities to select from. In the formalism
introduced by Challet and Zhang (1997), agents make their
choices among a set of tactics (or strategies as it is generally
m
called) whose size is 22 and, hence, may easily reach the
order of hundreds of thousands.
In our approach, the number of possible tactics is also
m
given by 22 . However, we do have a small set of personalities which are intended to represent the ordinary types of
commuters (or, more generally, all kinds of players in the
EFBP or by the MG formalisms). This set is not dependent
on m and only mimics the way people reason.
To date, we have set up nine different personalities which
are built up by one or more tactics. The advantage of the
treatment we give to the problem is twofold: we have a
small (and fixed) number of personalities, and the number
of tactics associated with them is (in general) much smaller
m
than the original 22 (which is not reasonable for human
reasoning).
Table 1 describes those nine personalities. Two of them
are memory-independent, which means that they represent
people who always do the same: either choose route A or
B no matter what has happened (personalities and , respectively). These are stubborn agents who stick to one tactics. According to surveys conducted among commuters,
the amount of people who behave like this is not small. For
instance, [14] report that 22% of people do not consider
changing routes, despite receiving traffic information.
and
Also in Table 1, we see that two personalities (
) actually need only a single memory position to represent their tactics. For a commuter of class (the persuasible ones), the route which had the smallest number of commuters (the “winner route”) the day before, will be chosen
today as a possible winner as well. On the contrary, agents
are the wayward agents: they do the opof personality
posite of what is suggested by the recent events as a good
choice. They select exactly the route which was chosen by
the majority of commuters the day before. In a real scenario, a commuter may take this decision possibly because
it is common sense that people are more likely to do what
was successful the day before.
Other interesting personalities are: the group of those
who choose route A or B if it was the winner in at least one
of the days kept in the history ( _ or _ respectively); in
all days in the history ( ^ or ^); and the random ( ) personality. The latter represents a commuter who either does
not care about the route taken, or a non-informed driver
(who will make a choice en route), or a person who does
not know the area well enough and hence is more likely to
follow any given instruction, or even those who just enjoy
varying routes so that commuting does not get boring.
A
History
DBYy Yz
0
0
1
0
0
1
1
1
B
W
P
m = 2. Table 3 describes the assignment of tactics to each
personality.
P
Pers.
A
B
A_
A^
B_
B^
P
W
R
B
A
B
R
Description
choose route A regardless of the history
choose route B regardless of the history
choose route A if it has won at least one day
choose route A if it has always won
choose route B if it has won at least one day
choose route B if it has always won
choice is the route that won the day before
choice is the route that lost the day before
choice of route is random
Table 1. Description of the Personalities.
Table 2 shows the history and the 16 possible tactics for
1
1
0
0
0
Tactics Number
2 3 4 5
0 1 0 1
1 1 0 0
0 0 1 1
0 0 0 0
6
0
1
1
0
7
1
1
1
0
(a) Tactics 0 to 7
History
DBYy Yz
0
0
1
0
0
1
1
1
W
A
0
0
0
0
0
8
0
0
0
1
9
1
0
0
1
Tactics Number
10 11 12 13
0
1
0
1
1
1
0
0
0
0
1
1
1
1
1
1
14
0
1
1
1
15
1
1
1
1
(b) Tactics 8 to 15
Table 2. History (left) — Possible Combinations of Winning Routes (for m=2) with Route
A Being the Winning Side (1) or Not (0)
Tactics (right) — Choosing Route A (1) or
Not (0), for m=2
y
Day Before Yesterday
z
Yesterday
The next section describes some simulations carried out
where n commuters have to simply select one of the two
routes (A and B). For our approach, it is important that the
society be as heterogeneous as possible. In this scenario,
there is no point in replacing personalities by others which
have performed better in the near past (e.g. through the use
of an evolutionary technique). Hence, our main interest is to
study the behaviour of all the types present in the population
and letting them compete in such scenario.
5. Result Analysis
Initially, we have tried to compare our results with those
obtained by Challet and Zhang (1997) regarding the fluctuation of route choice (in their case, bar attendance). They
have examined this attendance for n = 1001, for several
memory sizes, and the results show that the attendance indeed fluctuates around 50% for each bar (as depicted in
Figure 1(a)). We were able to draw the same conclusion
when performing the simulations, in which the only personality playing is (to match the random choice of their
approach where there is no policy to choose tactics). In fact
R
A
B
A_
A^
B_
B^
P
W
R
Column Number
15
0
2–15
8–15
0–13
0,2,4,...,14
12
3
0–15
cept that the choice of route A is around 430 on average
(m = 2). Remember that personalities ^ and ^ have,
respectively, more and less chance of selecting route A.
The next step was to simulate heterogeneous societies. A
particular (and the most interesting) case is when the nine
types of agents compete among themselves (Figure 1(b)). In
this case,
collects the most points, while collects the
lowest number of points. Personalities , , and _ are
second best and have very similar performances, followed
and _; next is ^ and ^, both having
closely by
a performance in between that of _ and . Note that,
in graph 1(b), personalities with similar performances are
plotted only once for the sake of clarity.
A
W
Table 3. Columns Representing each Personality
they randomly draw S tactics (from the set of all possible
for a given m) for each player (before the game starts), and
these choose one out of S to play at each time. The actual
choice is given by the history. This way, they do not really
consider all the possible tactics at each time, but the subset S , which is drawn ad hoc. A further conclusion of their
experiments is that the higher the memory size, the smaller
the standard deviation around the average bar attendance.
They explain this by an increasing “intelligence” of players,
who would be able to better cope with each other when having longer memories. Our results (again, for the simulation
with n = 1001 agents of type ) also indicate a reduction
of the standard deviation on increasing m, but it is negligible. This may be explained by the fact that the conditions
of their simulation and ours are actually not the same. They
use a subset S of tactics while we use the whole set. Also,
we have repeated each simulation 500 times while they do
not inform how many repetitions they have performed in
their experiments.
After this comparison exercise, we concentrated on our
main purpose, namely to perform simulations of societies
with different composition of agents. A particular case is
when n agents of a unique type play together. For type
, the results are already discussed above. Simulation of
homogeneous societies of agents type
and
are of no
interest: if they all choose the same route, none receives a
single point and hence the winning route is always the one
they do not choose.
Similarly, all agents of type
and
make the same
choice (of course varying from run to run since this depends
on the history) and hence collect no point. On average (of
500 runs) route A is selected by half of the agents.
As for societies having agents of type ^ only, the results for different memory sizes are not quite the same. This
is clear since the probability that route A always win decreases with increasing memory size. For m = 2, the number of agents choosing route A is near 570 on average (500
repetitions, n = 1001). Similar results apply for ^, ex-
P
B
W
A
B
0
100
B
600
400
200
0
200
300
Time Steps
400
500
(a) Route A Choice for m=2 and n=1001 agents of
type R
35
W
A, B, Bv
R
Av
A^, B^
P
30
25
Points (x1000)
A
A
P
AB
A
B
A
P
800
R
R
R
B
1000
Choice of Route A
Personalities
20
15
10
5
0
0
100
200
300
Time Steps
400
500
(b) Points
Time for the 9 Personalities Simulation, for m=2 and n=901
Figure 1. Simulation Results
The fact that agents of class
W were doing better than
the others indicated us that there was a tendency in our personalities to repeat the winning route of the previous day.
Note, however, that the personalities were not designed with
that in mind; they were all ordinary commuting strategies
like, e.g., having a preference for route A. We then proceeded to analyse the result of the simulation and, also using
an ad hoc analytical method, we verified that agents of class
were doing better because in fact all personalities we
had come up with, except
and _ slightly, tended coincidentally to repeat the previous winning route (which is
a quite plausible behaviour in real commuters, should they
have the knowledge about winning routes). This made us
realise that, if we could extend the (probabilistic) concept
of Hamming distance, which we explain next, to our multitactics personalities, this would be quite useful a concept in
analysing our simulations.
In 1998, Challet and Zhang introduced a concept to measure the difference between two tactics, the so-called Hamming distance. We briefly review it here: every tactics is
a set of 2m bits, which belongs to a hypercube of dimension 2m , Hm . For all tactics s 2 Hm ; s(i) 2 f0; 1g; i =
0 : : : ; 2m are the components of s. Thus the distance between two strategies s1 and s2 is:
W
W
dm (s1 ; s2 ) =
X
2m
1
2m
i=1
1
j s1 (i)
ij
XX
i
s2 (i) j :
(1)
dm (ko ; lp ):
(2)
o=1 p=1
A
B
RR
4 Note
m
m
A
A
B
P
W
A_
A^
B^
B_
R
Avg.
0,0000
1,0000
0,5000
0,5000
0,4464
0,3750
0,6250
0,5536
0,5000
0,5000
B
1,0000
0,0000
0,5000
0,5000
0,5536
0,6250
0,3750
0,4464
0,5000
0,5000
P
0,5000
0,5000
0,0000
1,0000
0,4821
0,3750
0,3750
0,5179
0,5000
0,4722
W
0,5000
0,5000
1,0000
0,0000
0,5179
0,6250
0,6250
0,4821
0,5000
0,5278
(a) Single-Tactics Personalities
A
B
P
W
A_
A^
B^
B_
R
Avg.
A_
A^
0,4464
0,5536
0,4821
0,5179
0,4923
0,4821
0,5000
0,5077
0,5000
0,4980
0,3750
0,6250
0,3750
0,6250
0,4821
0,3750
0,5000
0,5179
0,5000
0,4861
B_
0,6250
0,3750
0,3750
0,6250
0,5000
0,5000
0,3750
0,5000
0,5000
0,4861
B^
0,5536
0,4464
0,5179
0,4821
0,5077
0,5179
0,5000
0,4923
0,5000
0,5020
R
0,5000
0,5000
0,5000
0,5000
0,5000
0,5000
0,5000
0,5000
0,5000
0,5000
(b) Multiple-Tactics Personalities
Table 4. Hamming Distance between Personalities (2) for m=2.
j
In other words, Dm is the average of dm over all possible
combinations of tactics4 . For single-tactics personalities 1
and 2 are equivalent. Since and are opposite personalities Dm ( ; ) = 1. But for multiple tactics personalities
there are significant differences. The distance from a personality to itself is no longer 0, e.g., Dm ( ; ) = 0:5. On
the other hand, the Hamming distance is still the probability that the decision of two personalities differ (in our nine
personalities simulation).
AB
W
B
The Hamming distance is the probability that the decisions of two tactics differ from each other. So if two tactics
are opposite, their Hamming distance is dm = 1 and by
definition dm (s; s) = 0. Since we have introduced the use
of personalities which are a set of tactics, the concept of
Hamming distance needs to be generalised. We propose the
following: let K and L be two multiple-tactics personalities
with i and j tactics respectively, i.e., K = fk1 ; : : : ; ki g and
L = fl1 ; : : : ; lj g with i; j 2 f1; : : : ; 2m g. Referring to (1)
the extended Hamming distance is defined in the following
way:
Dm (K; L) =
For a good performance, it is crucial for an agent to have
an average Hamming distance of more than 0:5 regarding
the other personalities because, this way, the probability
to be in the minority group increases. This then explains
performs better than the others. Table 4
why personality
shows Dm calculated for all nine personalities.
that D (K; L) = D (L; K ).
Besides the case above, which is the most interesting
of all, some other particular cases follow, in which societies have an interesting composition. We next simulated a
, and (these are all singlepopulation of types , ,
tactics), for 401 agents in total5 . For m = 2, 500 runs of
the simulation showed that there is no fluctuation and the
average choice of route A is 200 commuters in it. Also, the
four types perform equally well, collecting nearly the same
ABW
P
5 In all simulations of heterogeneous societies, we have included an additional agent of type so as to keep all personalities with the same number of agents (as we need an odd number of agents). Note that inserting
one agent of type does not introduce any discrepancies, as the Hamming
distance of to any other personality is 1=2.
R
R
R
number of points during the simulated time (100 steps).
A B
Then we simulated societies of agents type _, _, and
(100 agents in each) because (operationally) the former
two strategies are described by almost as many tactics as
the latter, despite the fact that the reasoning behind the two
former and the latter are very different: selecting routes at
random has nothing to do with selecting A or B if this was
the winning route in one of the days before. The conclusion
was that the performance of them are not much different
(when they are found together), and the choice of route A
is 150 in average (m = 2, 500 repetitions). Afterwards,
we simulated a competition between types
and , for
m = 2 and m = 3, both with and without the presence
of personality . In each case there were 100 agents per
type.
and perform nearly the same, and when ’s are
present, better than these (around 8% at the end of the simulation). The average route choice of A was always 50%
of the total number of agents. All results regarding route
choice as well as points collected are given in Table 5 (Table 5(a) shows the parameters for simulations (i) to (ix), and
Table 5(b) shows the results for those same simulations).
R
W
W
P
R
P
R
Hor.
500
100
100
100
500
500
500
500
150
m
i
ii
iii
iv
v
vi
vii
viii
ix
2,3,4
2,3
2
2
2
2
2
2,3
2,3
Avg. A
500
500
573
428
449
200
1500
1000
300
n
1001
1001
1001
1001
901
401
3003
2001
303
(a) Simulation Parameters
Points
( 1000)
A B P W A_ B_ A^ B^
i
ii
iii
iv
v
vi
vii
viii
ix
R
244
0
42
25
25
25
25
44
17
25
34
25
50
5
50
5
24
20
44
20
42
25
25
44
4.6
(b) Simulation Results
Table 5. Summary of Some Simulations for
Route A Choice and Points Collect by each
Personality (at the end of the simulation horizon).
6. Conclusion
Our approach gives the El Farol Bar Problem (EFBP)
and the Minority Game (MG) a new direction in order to
tackle real commuters who are not able to process a huge
number of tactics, given that it increases exponentially with
the size of the memory (as it has been the case in the past literature). Agents posses different personalities which determine their distinct behaviours. For instance, the approach
by [8] makes no distinction between classes of players.
With our simulations we have proved that the type of players does matter, otherwise all personalities would have always performed the same. Even in cases where this is true
(e.g. _ and _ against ), the reasoning behind these
three personalities is completely different and must be taken
into account. Our main result is that, according to the distribution of personalities in a given population, specific agents
may perform better than average.
Given the EFBP is turning a paradigm for the study of
complex systems, our approach might function as a coordination mechanism for agents in this kind of scenarios. We
anticipate the point (more or less obvious) that information might became an important commodity in EFBP-like
scenarios. In particular, in the one we have studied here,
namely a commuting scenario, the more an agent is able to
infer about other agents personalities, the better it can perform.
This brings us to the next step of this work: to evolve
population of agents personalities. This can be done in two
ways: either agents are given information about traffic situation, so that they might try to learn new behaviours to
adapt themselves to these traffic situations; or we design,
ourselves, societies of personalities with various compositions. Since it is clear that the wayward is not evolutionary
stable, the latter direction is a very interesting one: to which
extent does the wayward agent still perform better?
Regarding the former direction, the key is to set up different scenarios for commuters in which they: (i) get information from a broadcasting system (e.g. radio) about
the traffic condition regarding both routes, and (ii) besides
getting it, they also receive advice (like “take this or that
route”) as mentioned in [14]. In both cases, we will have
to set up new personalities as well as change the existing
ones to deal with commuters who compete with those who
are not informed or, receive information but do not consider
it, or do consider it but not an advice, or, finally, those who
take into account neither information nor advice.
By simulating these situations, we will be able to examine how useful an information and/or advice is, and to
which extent people receiving the same information will actually be better off not considering/following them, which is
actually of high interest for commercial traffic information
providers.
A
B
R
Finally, another important research direction is to model
commuters in a less reactive way, following the work of
[11]. In order to achieve a more cognitive representation of
drivers, we plan to use their mental states to study coordination at a higher (strategic) level.
Acknowledgements
We would like to thank M. Schreckenberg for first hinting us on the Minority Game, and for fruitful discussions.
This work was partially supported by CNPq and BMBF
(project SOCIAT and research fellowships).
References
[1] W. B. Arthur. Inductive reasoning and bounded rationality.
American Economics Review, 84:406, 1994.
[2] A. L. Bazzan. A game-theoretic approach to distributed
control of traffic signals. In V. Lesser and L. Gasser, editors, Proceedings of the First International Conference on
Multi-Agent Systems (ICMAS’95), page 441, Menlo Park,
CA, 1995. AAAI Press / MIT Press.
[3] A. L. Bazzan. An Evolutionary Game-Theoretic Approach
for Coordination of Traffic Signal Agents. PhD thesis, Universität Karlsruhe, 1997.
[4] A. L. Bazzan, R. H. Bordini, and J. A. Campbell. Moral sentiments in multi-agent systems. In J. P. Müller, M. P. Singh,
and A. S. Rao, editors, Intelligent Agents V, Proceedings of
ATAL-98, number 1555 in LNAI, pages 113–131, Heidelberg, 1999. Springer-Verlag.
[5] A. L. Bazzan, J. Wahle, and F. Klügl. Agents in traffic modelling — from reactive to social behaviour. In W. Burgard,
T. Christaller, and A. B. Cremers, editors, Advances in Artificial Intelligence (KI-99), number 1701 in LNAI, pages
303–306, Berlin, 1999. Springer-Verlag.
[6] R. H. Bordini. Contributions to an Anthropological Approach to the Cultural Adaptation of Migrant Agents. PhD
thesis, University of London, 1999.
[7] C. Castelfranchi, F. de Rosis, and R. Falcone. Social attitudes and personalities in agents. In Proceedings of the AAAI
Fall Symposium on Socially Intelligent Agents, Cambridge,
Massachusetts, 8–10 November 1997.
[8] D. Challet and Y.-C. Zhang. Emergence of cooperation and
organization in an evolutionary game. Physica A, 246:407–
418, 1997.
[9] D. Challet and Y.-C. Zhang. On the minority game: Analytical and numerical studies. Physica A, 256:514–532, 1998.
[10] R. D’hulst and G. Rodgers. The hamming distance in the
minority game. Physica A, 270:514–525, 1999.
[11] B. Edmonds. Gossip, sexual recombination and the el farol
bar: modelling the emergence of heterogeneity. Journal
of Artificial Societies and Social Simulation, 2(3), 1999.
http://www.soc.surrey.ac.uk/JASSS/2/3/2.html .
[12] N. F. Johnson and P. M. Hui. Self-organized segregation
within an evolving population. Physical Review Letters,
82(16):3360–3364, 1999.
<
>
[13] N. F. Johnson, S. Jarvis, R. Jonson, P. Cheung, Y. R. Kwong,
and P. M. Hui. Volatility and agent adaptability in a selforganizing market. Physica A, 258:230–236, 1998.
[14] P. Nijkamp, G. Pepping, and D. Banister. Telematics and
Transport Behaviour. Springer-Verlag, Berlin, 1996.
[15] M. Ridley. The Origins of Virtue. Viking Press, London,
1996.
[16] J. Wahle, A. L. Bazzan, F. Klügl, and M. Schreckenberg.
Anticipatory traffic forecast using multi-agent techniques. In
D. Helbing, R. Herrmann, M. Schreckenberg, and D. Wolf,
editors, Traffic and Granular Flow ’99, Heidelberg, 1999.
Springer-Verlag. In print.