A satisﬁcing agreements model Guido Boella Universit´ a degli Studi di Torino guido@di.unito.it Gabriella Pigozzi Universit´ e Paris Dauphine gabriella.pigozzi@dauphine.fr Marija Slavkovik University of Luxembourg marija.slavkovik@uni.lu Leendert van der Torre University of Luxembourg leon.vandertorre@uni.lu Abstract—Satisﬁcing, the concept proposed by Herbert Simon, as an approach to reaching agreements is little explored. We propose a model for satisﬁcing agreement reaching for an adaptive collaborative group of agents. The group consists of one human agent familiar with the problem and arbitrarily many artiﬁcial agents. Our model raises to the team level the recognition-primed decision model constructed in the ﬁeld of cognitive decision-making by using social choice for reaching group opinions. Index Terms—agreement reaching; satisﬁcing; judgment ag- gregation; I. I NTRODUCTION Agreements are essential to the problem in agent coor- dination. Negotiation, in its many forms as argumentation, auctions, bargaining etc., is seen as an essential technology for reaching agreements [1]. However, negotiation is not suited for cooperation in fast changing environment. Negotiation protocols require several rounds of exchanges between the agents before an agreement is reached. Consequently the situation can change while the agents are still negotiating about the situation. It is in dangerous environments that we would like to replace the on-site human teams with robots and drones. Agreement-reaching problems can be addressed trough traditional decision-making; see for example Chapter 1 of [2]. Decision-making is driven by the concept of rationality associated with the decision-maker. A rational agent acts in his own best interest, i.e., chooses, given his knowledge about the world, those options that are optimal in the sense that they maximize the agent’s expected utility. Optimizing is difﬁcult when the agents’ resources are limited. People are not good rationalizers, (pg.ix [3]), but they are able to coordinate successfully under time pressure even when all adequate information is not available, when their goals are unclear and the procedures they have to follow are poorly deﬁned. Can these skills of high adaptiveness be advanced to artiﬁcial agents? The answer to this question begets another question: how do people make decisions under time pressure, in dynamic conditions in uncertain environments? Gary A. Klein and his associates, studied how ﬁreﬁghter commanders make decisions under extreme time pressure, [4]. They found that, when a commander has prior experience with a problem, which is normally the case, he acts according to the recognition-primed decision (RPD) model summarized in Figure 1. Klein et al.found that the RPD model exempliﬁes Herbert Simons’s [5] notion of satisﬁcing since the observed comman- Fig. 1. The recognition-primed decision model (pg. 203, [4]). ders were looking for the ﬁrst workable course of action rather than trying to ﬁnd the best possible option. Although ﬁreﬁghters operate as teams, the RPD model is a model of a single agent - that of the commander who coordi- nates the team. The commander is the one who assesses the situation, generates possible courses of action, evaluates them, and gives orders for actions that should be implemented by the rest of the team. The coordination through the commander is applicable when the commander is on the ground together with the team that executes his orders. Otherwise, the commander is not be able to asses the situation, nor make evaluations by himself. To apply the RPD model to teams, e.g., when the agents on the ground are robots or drones that coordinate remotely with the commander, we need to raise it from an individual raised to team level. This is the question that we address here, how to raise the recognition-primed decision model to a team level for use in multi-agent systems? Particularly, how can we do so without relying on negotiation? We consider a mixed human-robot team in which there is one human, called initiator, which has a role similar to that of the ﬁreﬁghter’s commander. The rest of the agents are artiﬁcial, called executors. Unlike the commander, the initiator is not on the ground and has to fully rely on the executors for the following processes: situation assessment, verifying expectancies, and evaluating the potential course of action. The situation assessment is the process in which the initiator matches the problem with a possible solution (a goal). The goal is good enough if and only if a certain combination of cues can be identiﬁed as (not) present. Verifying expectancies is the process in which the agreements on cue’s presence or absence and goals are veriﬁed as the agents proceed with pursuing the goal. While the commander of the RPD model can identify the cues himself, the initiator needs to obtain an agreement on them, by considering the opinions of the executors. The initiator also needs the input from the executors for verifying expectancies. The reasoning according to the RPD model is very fast;