Łukasiewicz Games
Enrico Marchioni Michael Wooldridge Institut de Recherche en Dept of Computer Science Informatique de Toulouse University of Oxford, UK Université Paul Sabatier, France [email protected] [email protected]
ABSTRACT ual goal 'i, expressed as a formula of propositional logic. Strate- Boolean games provide a simple, compact, and theoretically attrac- gic concerns in Boolean games arise from the fact that whether i’s tive abstract model for studying multi-agent interactions in settings goal is in fact satisfied will depend in part on the choices made where players will act strategically in an attempt to achieve per- by other players. Players in Boolean games can be understood as sonal goals. A standard critique of Boolean games, however, is representing non-deterministic computer programs, and the over- that the binary nature of goals (satisfied or unsatisfied) inevitably all framework of Boolean games provides an elegant mathematical trivialises the nature of such strategic interactions: a player is as- model through which to investigate issues of strategic interaction sumed to be indifferent between all outcomes that satisfy his goal, in multi-agent systems. and indifferent between all outcomes that do not satisfy his goal. Now, a standard critique of Boolean games is that the binary na- In this paper, we introduce Łukasiewicz Games, which overcome ture of goals (satisfied or unsatisfied) inevitably trivialises the na- this limitation by considering goals to be specified using finitely- ture of strategic interactions. For example, players are assumed to valued Łukasiewicz logics. The significance of this is that formu- be indifferent between all outcomes that satisfy their goal, and are lae of Łukasiewicz logic can express every continuous piecewise indifferent between all outcomes that do not satisfy their goal. This linear polynomial function with integer coefficients over [0, 1]n, assumption is clearly a gross simplification for many situations, thereby allowing goal formulae in Łukasiewicz Games to naturally a concern which led researchers to extend the original Boolean express a much richer range of utility functions. After introducing games model with costs, leading to richer and more realistic pref- the formal framework of Łukasiewicz Games, we present a num- erence structures for agents [11]. While these refinements make it ber of detailed worked examples to illustrate the framework, and possible to model much richer types of interaction, the inherently then investigate some of the properties of Łukasiewicz Games. In dichotomous nature of preferences in Boolean games is surely one particular, after investigating the complexity of decision problems of their most debated features, and the work in the present paper is in Łukasiewicz Games, we give a logical characterisation of the directly motivated by this limitation. existence of Nash equilibria in such games. Specifically, we introduce Łukasiewicz Games, which overcome this limitation of Boolean games by allowing goals to be specified as formulae of finitely-valued Łukasiewicz logics [3]1. The ratio- Categories and Subject Descriptors nale for using Łukasiewicz logic in this way is given by the Mc- I.2.11 [Distributed Artificial Intelligence]: Multiagent Systems Naughton Theorem, which says that every continuous piecewise I.2.4 [Knowledge representation formalisms and methods] linear polynomial function with integer coefficients over [0, 1]n can be expressed as a formula of Łukasiewicz logic in n variables (see, General Terms e.g., [3]). Thus, Łukasiewicz logic provides a natural, compact, for- mally well-defined and expressive logical representation language Theory for payoff functions, allowing much richer preference structures than is easily possible in conventional Boolean games. Keywords The present paper makes three main contributions. First, we Boolean games, Łukasiewicz Games, Łukasiewicz logic provide a formal definition of Łukasiewicz Games. Second, we present a number of detailed worked examples, which illustrate how a range of strategic scenarios can be formalized within this 1. INTRODUCTION framework. In particular, we argue, these scenarios cannot natu- Boolean games provide a simple, compact, and theoretically attrac- rally be formalized using conventional Boolean games. Third, we tive abstract model for studying multi-agent interactions in settings investigate properties of Łukasiewicz Games. We show that despite where players will act strategically in an attempt to achieve private their expressive power, the key decision problems for Łukasiewicz goals [8, 2, 5, 6]. In a Boolean game, each player i exercises unique Games are no more complex than for conventional Boolean games; control over a set of Boolean variables Vi, and will attempt to as- in addition, we give a logical characterisation for the existence of sign values for these variables in such a way as to satisfy an individ- equilibria in Łukasiewicz Games.
Appears in: Alessio Lomuscio, Paul Scerri, Ana Bazzan, and 1 Michael Huhns (eds.), Proceedings of the 13th International Con- In the remainder of the paper, when we refer to Łukasiewicz logic, ference on Autonomous Agents and Multiagent Systems (AAMAS it should be understood that we are referring specifically to finitely- 2014), May 5-9, 2014, Paris, France. valued Łukasiewicz logics, as distinct from the many other exten- Copyright c 2014, International Foundation for Autonomous Agents and sions of Łukasiewicz logic that have been investigated in the liter- Multiagent Systems (www.ifaamas.org). All rights reserved. ature [3].
837 2. PRELIMINARY DEFINITIONS 3. ŁUKASIEWICZ GAMES Since Łukasiewicz logic is fundamental to our present work, but We now introduce the framework of Łukasiewicz games. First, let is not as widely known as the classical logics that underpin con- V = p1,...,pm be a finite set of propositional variables, as ventional Boolean games, we begin by introducing the concepts of above.{ Our games} are populated by a finite set of players P = Łukasiewicz logic that are used in the remainder of the paper. P1,...,Pn (also referred to as “agents”). Note that throughout { } The language of Łukasiewicz logic Ł[3] is built from a countable this paper, we assume that P = n. Each player Pi controls a | | set of variables V = p1,p2,... , the binary connective “ ”, and subset of propositional variables Vi V, so that the sets Vi form a { } ! ✓ the truth constant 0 (for falsity). Further connectives are defined as partition of V. The fact that player Pi is in control of the set Vi of follows: propositional variables means that Pi has the unique ability within the game to choose values for the variables in Vi. It is assumed that ' is ' 0¯, ' is '&(' ), 1 k 1 ¬ ! ^ ! variables take values from the set Lk = 0, ,..., , 1 . '& is (' ), ' is ((' ) ), { k k } ¬ !¬ _ ! ! A strategy for an agent Pi is a function si : Vi Lk, which ' is ( '& ), ' is (' )&( '), ! ' is '¬&¬ , ¬ d('$, ) is ('! ). ! corresponds to a valuation of the variables controlled by Pi.A ¬ ¬ $ strategy profile is a collection of strategies (s1,...,sn), one for We ofter write n' as an abbreviation for ' ', with n 1. each player. Every strategy profile directly corresponds to a valua- ··· tion function e : V Lk and vice versa; we find it convenient to n ! Let Form denote the set of formulae of Łukasiewicz logic. A abuse notation a little by treating strategy profiles as valuations and valuation, e, is a mapping e : V [0|, 1],{z which} assigns to all valuations as strategy profiles. ! propositional variables a value from the real unit interval; we re- As in conventional Boolean games, we assume that each player e(0) = 0 c quire that . The semantics of Łukasiewicz logic are then is associated with a Łk-formula 'i, with propositional variables defined, with a small abuse of notation, by extending the valuation from V, whose valuation is interpreted as the payoff function for e to complex formulae. Although strictly speaking we only need state the rule for the “ ” operator (as we can define the remaining player Pi. That is, the player Pi seeks a valuation e that maximises ! the value of the corresponding function f' . Of course, not all the operators in terms of this operator and 0), we present the complete i ruleset in the interest of clarity: variables in 'i will in general be under Pi’s control and, conse- quently, the utility Pi obtains by playing a certain strategy (i.e., e(' )=min(1 e(')+e( ), 1) choosing a certain variable assignment) also potentially depends in e!( ')=1 e(' ) part on the choices made by other players. e('&¬ )=max(0 ,e(')+e( ) 1) Łukasiewicz Games can be seen as a generalization of Boolean e(' )=min(1,e(')+e( )) Games [8, 2], since the latter are obviously a special case of the e(' )=max(0,e(') e( )) e(' )=min(e('),e( )) former. However, Łukasiewicz Games also incorporate the notion e(' ^ )=max(e('),e( )) of cost/effort, given that each player’s strategic choice can be seen e(' _ )=1 e(') e( ) as an assignment to each controlled variable carrying an intrinsic e(d('$, )) = e ('|) e( ) | cost. This makes it possible to formalize situations in which agents | | aim at a better tradeoff between the costs of making certain choices A valuation e is a model for a formula ' if e(')=1. A valuation and the resulting payoff. e is a model for a theory T , if e( )=1, for every T . XAMPLE Consider the following example with two players 2 E 1. In this paper, we restrict our attention to finite-valued Łukasiewicz P1 and P2, who control variables x and y, z , respectively. The { } { } logics Łk. In such logics, it is assumed that the domain is a set of payoffs are given by the following formulae: 1 k 1 the following form: Lk = 0, k ,..., k , 1 . The notions of { } '1 := x y z and '2 := x z. valuation and model for Łk are defined analogously just replacing ^¬ ^ [0, 1] by Lk as set of truth values for the logic, that is, valuations In order for P2 to maximize the payoff, it suffices to assign s2(z)= are functions with the signature e : V Lk. It is sometimes use- 1, no matter what P does. Player P , however, has no chance of ! 1 1 ful to introduce constants in addition to 0 that will denote values in maximizing the payoff except by assigning s1(x)=1and hoping c the domain Lk. Specifically, we will denote by Łk the Łukasiewicz that s2(z)=1and s2(y)=0. The latter makes sense, since, logic obtained by adding constants c for every value c Lk.We given s (z)=1, any assignment to y does not change the value 2 2 assume that valuation functions e interpret such constants in the of '2, so P2 has no incentive in assigning any other value to y but natural way: e(c)=c. 0 and raising the costs. Still, P2 knows that P1 needs s1(x)=1 A well-known result by McNaughton established that every con- and consequently can get the maximum payoff by simply keeping tinuous piecewise linear polynomial function with integer coeffi- the costs at minimum with s2(z)=0and s2(y)=0. cients over [0, 1]n (called a McNaughton function) is definable by a formula in Łukasiewicz logic [3]. We now formally define Łukasiewicz games. c In the case of finite-valued Łukasiewicz logics Łk, the functions DEFINITION 1(ŁUKASIEWICZ GAMES). AŁukasiewicz game c defined by a formula are combinations of the restrictions of Mc- on Łk is given by a structure: n Naughton functions over (Lk) and the constant functions for each G c = P, V, Vi , Si , 'i c Lk. Notice that the class of functions definable by Łk-formulas G h { } { } { }i 2 n exactly coincides with the class of all functions f :(Lk) Lk, where: for every n 0.2 In this sense, we can associate to every formula! c n 1. P = P1,...,Pn is a finite set of players. '(p1,...,pn) from Łk a function f' :(Lk) Lk. { } ! Finally, note that the satisfiability problem for finite Łukasiewicz 2. V = p1,...,pm is a finite set of propositional variables logics is no more complex than for classical propositional logic: the taking{ values from} problem is NP-complete [7, 3]. 1 k 1 Lk = 0, ,..., , 1 . 2 k k This result can be obtained by using Lemma 3. ⇢
838 3. Vi V is the set of propositional variables under control of The NON-EMPTINESS problem is the problem of determining for a ✓ player Pi, so that the sets Vi form a partition of V, i.e.: given game whether there exist any pure strategy Nash equilibria G p n for the game. The problems are co-NP-complete and ⌃2-complete V = Vi, and for i = j, Vi Vj = . for conventional Boolean games. And, perhaps surprisingly, we 6 \ ; i=1 have: [
4. Si is the strategy set for player i that includes all valuations THEOREM 1. The MEMBERSHIP and NON-EMPTINESS prob- p si : Vi Lk of the propositional variables in Vi, i.e. lems for Łukasiewicz games are co-NP-complete and ⌃2-complete, ! respectively. Si = si si : Vi Lk . { | ! } The proof of this can be adapted from proofs of the corresponding c 5. 'i is a Łk-formula, built from variables in V, whose associ- results for conventional Boolean games: the key point is that strate- ated function gies for Łukasiewicz games are “small witnesses” in the terminol- t ogy of computational complexity. A strategy for a player simply f'i :(Lk) Lk ! consists of a value from the domain Lk for every variable con- corresponds to the payoff function (also called utility func- trolled by that player. Since values in Lk are rational numbers that tion) of Pi, and whose value will be determined by the valu- can be assumed to be expressed in binary, they have a compact rep- ations in S1,...,Sn . resentation. { }
A strategy profile ~s for is a tuple ~s =(s1,...,sn), with each G 4. EXAMPLES si Si being the strategy selection for the corresponding player in 2 . Given a strategy si for Pi, we denote by s i the collection of We now introduce a number of examples to illustrate Łukasiewicz G strategies (s1,...,si 1,si+1,...,sn) not including si, and S i is games, and in particular, we choose examples that we believe can- the set of all s i’s. With an abuse of notation, we use not be easily expressed in the framework of conventional Boolean games (i.e., using classical logic to express goals). f'i (s1,...,sn) and f'i (si,s i) 4.1 Generalized Matching Pennies to denote Pi’s payoff under the strategy profile (s1,...,sn): re- The first game we present is a generalization of Matching Pennies3, call that 'i defines a payoff function f' , and a strategy profile i the classic example of a zero-sum game without a pure strategy (s1,...,sn) corresponds to a valuation e : V Lk. We now introduce the concepts of dominance,! best response and equilibrium. In the original game, two players P1 and P2 both Nash equilibrium adapted to our framework. In what follows, we have a penny and must secretly choose whether to turn it to head always assume we are working with an arbitrary n-player Łukasiewicz or tails, revealing their choices simultaneously. If their choices are game on Łc . the same, then P1 takes both pennies; if they are different, P2 takes G k both. DEFINITION 2(DOMINANCE). Let be a Łukasiewicz game. Now, imagine that both players must perform an action with a G A strategy si Si for player Pi is called strictly dominated (weakly certain cost and are in charge of the variables p1 and p2, respec- 2 dominated) if there exists a strategy si0 Si such that for all tively. P1’s overall strategy is to be as close as possible to P2’s s i S i 2 2 choice. In contrast, P2 wants to keep the greatest possible distance between the choices. The players’ strategy spaces are given by the f'i (si,s i)