lake Tax as a Automated

When autonomous agents attempt to coordinate action, it is often necessary that they reach some kind of consensus. Reaching such a consensus has traditionally been dealt with in the Distributed Artific.ial Intelligence literature via the mechanism of negotiation. Another alternative is to have agents bypass negotiation by using a voting mechanism; each agent expresses its preferences, and a group choice mechanism is used to select the result. Some choice mechanisms are better than others, and ideally we would like one that. cannot be manipulated by an untruthful agent. One such non-manipulable choice mechanism is the Clarke tax [Clarke, 19711. Though theoretically attractive, the Clarke tax presents a number of difficulties when one attempts to use it in a practical implementation. This paper examines how the Clarke tax could be used a.s an effective “preference revealer” in the domain of automated agents, reducing the need for explicit negotiation. ackground and Motivation When autonomous agents attempt to coordina.te action, it is often necessary that they reach some kind of consensus. Multi-agent activity is obviously facilitated by, and sometimes requires, agreement by the agents as to how they will act in the world. Reaching such a. consensus has traditionally been dealt with in the Distributed Artificial Intelligence literature via the mechanism of negotiation [Rosenschein and Genesereth, 1985; Durfee, 1988; Sycara, 1988; Kuwabara and Lesser, 1989; Conry et al., 1988; Kreifelts and von Martial, 1990; Kraus and Wilkenfeld, 1990; Laasri et al., 19901. One scenario [Zlotkin and Rosenschein, 1990b] that has been addressed in the research on negotiation involves a group of agents and a negotiation set. The role of negotiation is to reach consensus by allowing the agents to choose one element of this set. The main concern of a negotiation protocol is usually that the agreed-upon decision will be optimal in some sense. A basic assumption of the negotiation process is that each of the pa.rticipating a.gents has a private preference relation over the set of alternatives. Optimality is measured with respect to these preferences. Given the agents’ preferences and the optimality criterion, determining the optimal choice is a matter of direct computation. Thus, the substantive role of the negotiation process is to reveal preferences. If there existed another method of revealing the true preferences of agents, the need for negotiation would be 1essened.l There have been several attempts, both inside of Artificial Intelligence (AI) and outside, to consider market mechanisms as a way of revealing agents’ true preferences (and thus efficiently allocate resources). Notable among the AI work is that of Smith’s Contract Net [Smith, 19781, Malone’s Enterprise system [Malone et al., 19881, and the work of Miller and Drexler on Agoric Open Systems [Miller and Drexler, 19881. Following in this line of work, we present an alternative method for revealing agents’ preferences, the Clarke tax, and consider how it could be used among automated agents. The General Framework Assume a group of N a.gents A operating in a world currently in the state SO, facing the decision of what to do next. One way of formulating this problem is to consider that the a.gents are trying to agree into which member of the set S of m. possible states the current world should be moved. Each agent in A has a worth, or utility, that he associates with each state; that worth gives rise to a preference relation over states. Agent i’s true worth for state k will be denoted by W(i, k). However, the preferences declared by an agent might differ from his true preferences. The decision procedure that chooses one state from S is a function from the agents’ declared preferences to a member of the set { 1, . . . . m}. It maps the agents’ declared preferences into a group decision as to how the world will be transformed. ‘This assumes the agents’ preference relations are static during the negotiation process. Otherwise, the negotiation itself could cause the agents to acquire new information and alter their preferences, thus remaining useful. EPHRATI & ROSENSCHEIN 173 There a.re many decision procedures that reach a pareto optimal decision, but they suffer from two major drawbacks. First, they are manipulable, which means that an agent can benefit by declaring a preference other than his true preference.2 Second, they only take into consideration the ordinal preferences of the agents, i.e., the order in which an agent ranks choices, without assigning relative weights. Attempts to overcome this latter drawback motivated the development of voting procedures based on cardinal orderings over alternatives (that is, allowing agents to weight their choices, including negative weights). The most straightforward procedure, “sealed bidding,” allows each voter to specify an amount of money (positive or negative) for each alterna,tive. The alternative that has the maximal sum is chosen. Positive bids are then collected, and some of this money is then handed over to those agents (if any) who gave negative bids with respect to the chosen alternative. Although a voter can guarantee his max-min value [Dubins, 19771, he does have an incentive to underbid-if he assumes other agents will cause some alternative to win even without the full strength of his vote, he can underbid, get what he wants, and pay less. However, since the a.gent might be mistaken as to how others will vote, a sub-optimal alternative might be chosen. In the literature of Economics, this problem is known as the free rider problem; for many years it was believed to be unsolvable. A solution to the problem was presented by E. H. Clarke in 1971 [Clarke, 1971; Clarke, 1972; Straffin, 19801. In the following sections, we present Clarke’s scheme and analyze ways in which it can be used by communities of automated agents.