Abstract. While both games and Multi-Objective
Optimization (MOO) have been studied extensively in the
literature, Multi-Objective Games (MOGs) have received less
research attention. Existing studies deal mainly with
mathematical formulations of the optimum. However, a
definition and search for the representation of the optimal set,
in the multi objective space, has not been attended. More
specifically, a Pareto front for MOGs has not been defined or
searched for in a concise way. In this paper we define such a
front and propose a set-based multi-objective evolutionary
algorithm to search for it. The resulting front, which is shown
to be a layer rather than a clear-cut front, may support players
in making strategic decisions during MOGs. Two examples are
used to demonstrate the applicability of the algorithm. The
results show that artificial intelligence may help solve
complicated MOGs, thus highlighting a new and exciting
research direction.
I. INTRODUCTION
In conventional game theory problems, the usual assumption
is that decision makers (players) make their decisions based
on a scalar payoff. But in many practical problems in
economics and engineering, decision makers must cope with
multiple objectives or payoffs. In such problems, a vector of
objective functions must be considered. A MOG may be
defined as a problem in which one player aims to minimize a
set of objectives, while the other player aims at maximizing
these objectives. As an example of such a game, consider a
player who possesses a set of boats. These boats are to set
sail from several possible docking locations, so that one of
them will reach a target point in the shortest time possible.
While sailing, the boats must escape an opponent vessel that
is guarding the target and is capable of traveling twice as fast
as the boats. When the vessel intercepts a boat, that boat is
eliminated from the game. The MOG is defined such that the
first opponent seeks to minimize the distance traveled by one
of its boats (minimizing the time taken to get to the target)
while also minimizing its loss of boats (sacrificed for the
sake of optimization). On the other hand, the opponent's
objectives are to maximize the number of intercepted boats
and to maximize the time (distance) taken for any of the
A. Avigad is with the Mechanical Engineering Department at Ort
Braude College of Engineering, Karmiel, Israel (phone: 97249901767; fax:
97249901866; e-mail: gideona@braude.ac.il).
E. M. Eisenstadt is with the Mechanical Engineering Department at Ort
Braude College of Engineering, Karmiel, Israel (e-mail:
erella@braude.co.il).
M. C. Weiss is with the Computer Science Department at Ort Braude
College of Engineering, Karmiel, Israel (e-mail: miri@braude.co.il).
opponent's boats to reach the target point. The strategies
adopted by the first player may include, but are not limited
to, where the boats should start from and how many and
which boats should be sacrificed. The strategies used by the
vessel may include where to start, when to start and whether
to intercept the boat closest to the target or the nearest boat.
As in the case of single objective games, the notion of
equilibrium can be defined in terms of unfruitful deviation,
from the equilibrium strategies. According to [1], in the
MOG setting this can be interpreted as follows:
Deviations from the equilibrium strategies do not offer any
gains to any of the pay-off functions for any of the players.
All existing studies related to MOGs refer to definitions for
this equilibrium (e.g., [2]). The predominant definition of
this equilibrium, termed the Pareto-Nash equilibrium, has
been suggested in [3]. The Pareto-Nash equilibrium uses the
concept of cooperative games, because according to this
notion, sub-players under the same coalitions should,
according to the Pareto notion, optimize their vector
functions on a set of strategies. On the other hand, this notion
also takes into account the concept of non-cooperative
games, because coalitions interact on the set of situations and
are interested in preserving the Nash equilibrium between
coalitions. An extension to the work in [3] may be found in
[1], where several algorithms have been suggested for
searching for these equilibrium strategies.
MOGs are commonly solved by linear programming, e.g.
[4], in which a multiple objective zero-sum game is solved.
A non-zero-sum version of a MOG has been solved by
Contini [5]. Detailed algorithms for finding Pareto Nash
equilibrium-related strategies may be found in [1] and [3]. In
all of these cases, weights are altered in order to search for
one equilibrium solution at a time. This means the algorithms
must be executed sequentially in order to reveal more
equilibrium points. Artificial intelligence-based approaches
have been applied to MOGs within the framework of "fuzzy
multi-objective games." In such studies (e.g., [6]) the
objectives are aggregated to a surrogate objective in which
the weights (describing the players' preferences towards the
objectives) are modeled through fuzzy functions.
II. BACKGROUND
A. MOO and EMO
The topic of MOO concerns the search for solutions to
many real world problems, dubbed appropriately enough
Multi-objective Problems (MOPs). In cases of contradicting
objectives, there is no universally accepted definition of an
Optimal Strategies for Multi Objective Games and Their Search by
Evolutionary Multi Objective Optimization
G. Avigad, E. Eisenstadt, M. Weiss Cohen
978-1-4577-0011-8/11/$26.00 ©2011 IEEE 166