The experimental study of human strategic behavior over networks is a topic of great current interest in the literature. The work by Kearns and others on network coloring and consensus games [6]–[10] has been particularly influential. Judd et al. [17] investigated how subjects choose between playing either a dominant or a submissive role in a network game, documenting the importance of fairness. Kearns et al. [18] performed experiments on network formation games when there is a cost for creating links. Suri and Watts [19] conducted experiments in which individuals connected over networks play local public good games. Wang et al. [20] studied multi-player prisoner's dilemma games in which subjects can propose and delete links to other players, showing that partner selection increases cooperation. Brautbar and Kearns [21] introduced a network formation game in which players need to maximize their clustering coefficients. Compared to these previous works, we focus on isolating behavioral principles of human interaction (in the context of maximum matching games) and using these principles to formulate algorithmic predictions of outcomes.

As social interaction naturally induces strategic behavior, our work is also closely related to game theory. Indeed, several authors proposed game theoretical models of human interaction over social networks. Topics vary from diffusion and contagion over networks [16], [22], [23] to strategic information retrieval [24], [25], models of segregation [26] and bargaining over networks [27], to mention a few. The main element that distinguishes our work from the game theory literature is that we focus on the algorithmic processes involved in strategic thinking and the ensuing collective dynamics rather than on equilibria. Moreover, our algorithmic model is motivated and supported by experimental data.

Finally, matching theory has received notable attention throughout the decades, both in the context of game theory and economics [28]–[31], and in the development of algorithms for the maximum matching problem [32]–[36]. We point out that our simplified setup constitutes a simplification of the richness and heterogeneity of the ties in real social networks, as the subjects have no preference over each other, all the ties are equivalent, and interaction has no cost. However, such a simplified model leads to a tractable analysis and to the formulation of a general principle of collective behavior.

Before formulating our algorithmic model, we conducted four sessions of experiments, each with a different pool of sixteen undergraduate students connected over a virtual network. Subsequently, to validate our model, we ran an additional session of experiments with a pool of twenty four subjects on a set of networks that included small world and preferential attachment networks. In each of the first four sessions the subjects were asked to solve the matching game on a pool of over 70 networks. All networks admitted a perfect matching, namely a maximum matching with no unmatched nodes. We considered networks classified into four groups: bipartite networks admitting a unique perfect matching, bipartite networks admitting multiple perfect matchings, non-bipartite networks admitting a unique perfect matching, non-bipartite networks admitting multiple perfect matchings. Within each group, networks were randomly generated. As a remark, a bipartite network is a network whose nodes can be divided into two disjoint sets and such that every edge connects a vertex in to one in . If this property does not hold, we say that the network is non-bipartite. Subjects sat in front of workstations for the entire two-hour duration of the session and had no eye-contact with each other. For each matching game, a network was chosen, subjects were randomly assigned to its nodes, and each subject interacted with its neighbors by making or accepting proposals to form matched pairs using the interface shown in Figure 1. Each subject could control the node in the center of the screen and could only see its neighbors and, among those, distinguish which of them were currently matched (marked in dark green). A subject could make proposals or accept proposals by selecting a neighbor with a mouse click, and could only have one outstanding proposal at a time to form a matched pair (circled in yellow). While subjects knew whether a neighbor is matched or unmatched, they did not have direct knowledge of any outstanding requests to their neighbors other than their own. If two neighbors selected each other, a pair was formed (a bright green link appeared between them) which could be broken when one of the partners selected another neighbor. As a remark, since a pair was formed when two subjects selected each other and each subject could make a single selection at a time, each subject could be part of a single pair at a time (with one of its neighbors).

If a perfect matching was found within the time limit of five minutes, the game was declared solved and each participant was rewarded by $.50 or $1 depending on the session, otherwise the game ended with no reward. The number of games in an experimental session was not fixed, but games were run for the two-hour duration of the session. Therefore, the number of games and the cumulative reward in a session depended on the performance of the participants, providing an additional incentive to coordinate.

The networks used in this first set of experiments can be divided into four classes: bipartite, non bipartite, unique perfect matching, multiple perfect matchings. Two one-tailed Welch's t-tests confirmed the hypotheses that it is harder for humans to complete the matching game on non-bipartite than on bipartite networks (-value ); and that non-bipartite networks with unique perfect matching are more difficult to solve than non-bipartite networks with multiple perfect matchings (-value ). No statistically significant difference was found between the completion time of bipartite networks with unique and with multiple perfect matchings. We believe that this depended on the small network size of sixteen nodes and we did not explore larger bipartite networks further.

One of the main behavioral properties that emerged from the experimental data is that matched players may break their current matching only if they have been requested by an unmatched neighbor. In particular, in of the games no player ever violated this rule at any time during the game. In the remaining games, over percent of the moves were in agreement with this rule. Therefore, this property led to the following modeling assumption:

Assumption 1 (Prudence) A matched node does not break its current matched pair if it does not receive any request from other neighbors.

Two remarks are in order. First, note that this behavioral rule is peculiar to the matching problem since each player needs to choose a partner but also needs to be chosen. Second, notice that a matched subject with unmatched neighbors has some incentive to behave non-prudently and break the current match, because the subject can infer from the status of its neighbors that the perfect matching is not reached yet. However, experimental data shows that this rarely happens.

For each node , let indicate 's current preference. In other words, is the unique node to which has currently proposed to. will be null if does not have a current proposal. If two neighbors and currently prefer each other (i.e., and ), then consider them matched and the edge as part of the matching. Assume that each node knows if a neighbor is matched or unmatched.

Given the prudence property, we model the algorithmic behavior of humans using the algorithm shown in Table 1. The algorithm is specified by the implementation of two functions, called and , which are placeholders for the behavior that node would follow depending on whether is matched or unmatched. We consider a synchronous setting, in which time is divided into rounds, and at the beginning of each round each node observes its status and the status of its neighborhood and then decides on an action to take.

In the following we provide a canonical implementation of the functions and consistent with the prudence property. does not change the current value of with probability , while with probability accepts the proposal from a neighbor uniformly at random from among the neighbors with if any; if there is no neighbor with , then it proposes to a node uniformly at random from among the unmatched neighbors if any; otherwise it proposes to a node uniformly at random from among all the matched neighbors. In other words, unmatched nodes prefer neighbors who requested them over other unmatched neighbors, and unmatched neighbors over matched neighbors. As for matched nodes, accepts a proposal from a neighbor uniformly at random from among the neighbors with (note that 's current partner is one of them). We remark that the simulations' performance and the fit with the experimental data was practically insensitive to the value of chosen in the run of the algorithm.

In this section we present our analytical results regarding the convergence behavior of the Prudence algorithm. In particular, our results describe how well the algorithm performs in finding a large matching and the time it takes in terms of the number of rounds required. Due to space constraints, we only present proof sketches here. Complete details of the proofs are deferred to the SI.

We define a matching at round as the set of matched edges at the beginning of round of the algorithm. We first claim that the prudence property implies that the size of the matching does not decrease with time. The proof is immediate and it is omitted.

Claim 1 The size of the matching at round is non-decreasing as increases.

We then observe that the behavior of the Prudence algorithm can be described by a Markov chain over matchings. A transition from a matching to a matching is made by selecting an edge such that at least one among and is unmatched, and setting if are both unmatched, and if exactly one of and is matched in and is the matching edge. This Markov chain is reversible when restricted to matchings of the same size. Since the Markov chain is memory-less and has positive probability of reaching a maximum matching, we conclude that the Prudence algorithm enjoys self-stabilization.

Claim 2 The Prudence algorithm is a self-stabilizing algorithm.

Our first theorem says that a -approximate matching will be reached quickly in networks with bounded degree.

Theorem 1 In any bounded-degree graph on nodes, the expected number of rounds for the algorithm to reach a -approximate matching is .

The key idea of the proof is to show that, in expectation, the “distance” in terms of number of matched pairs to the smallest maximal matching shrinks by a constant factor in each round of the Prudence algorithm. Since it is well known that any maximal matching is a -approximation of the maximum matching, the result then follows.

We remark that the assumption of having bounded degrees is necessary as there are unbounded degree graphs in which a polynomial number of rounds is required with high probability to achieve a -approximation. However, in this case, a polynomial number of rounds is also enough to achieve any constant approximation: indeed, as the next theorem states, the Prudence algorithm provides a PTAS (polynomial time approximation scheme) for the maximum matching problem. Given a graph , denotes its maximum degree.

Theorem 2 For any graph of nodes, and , the Prudence algorithm reaches a -approximate matching in rounds with probability at least .

The theorem implies that, for any constant , a matching whose size is within a fraction of the size of the maximum matching is reached in polynomial time. For bounded-degree graphs, this result also holds for , implying that in this case a maximum matching can be reached in polynomial time.

To prove the theorem, we track the progress of the algorithm towards an approximate maximum matching, using the concept of an augmenting path. An augmenting path is a path of odd length which alternates between matched and unmatched edges and whose extreme edges are unmatched. It turns out that there is a close connection between the size of a shortest augmenting path in a matching and how close the matching size is to the size of a maximum matching. More specifically, we use the following lemma due to Hopcroft and Karp [32].

Lemma 1 Consider any matching that does not admit augmenting paths of odd length or smaller. Then, the size of is at least a fraction of the size of a maximum matching.

Hence, to prove Theorem 2, we need to show that short augmenting paths (for a suitably chosen ) are solved in a short amount of time. It is useful to consider a particle analogy to understand the process that eliminates short augmenting paths. We consider each unmatched node as a particle. Particles move around the graph from node to node as nodes change their status between matched and unmatched states dictated by the random choices in the algorithm. There are exactly two particles along an augmenting path, situated at the extreme nodes. To understand how an augmenting path gets shorter and eventually vanishes, we consider how the two particles move closer to each other along the path.

Let denote a shortest augmenting path. If the extreme unmatched node proposes to and accepts the proposal breaking the current match with , then the particle moves from to . A similar argument applies to the other end of the path. Also, the minimality of the path guarantees that the internal nodes do not change their current matching as they have no unmatched neighbor. It follows that the particles become closer to each other and the augmenting path gets shorter. Using this approach, we can prove that with suitable probability the length of the shortest augmenting path shrinks after each round. When an augmenting path becomes an edge (that is, a path of length one), and if the extreme unmatched nodes select each other as partners, the particles and the path vanish, yielding an increment to the size of the matching. Hence, a key step of our proof is to lower bound the probability that an augmenting path of length vanishes, and then to apply Lemma 1 to relate the existing augmenting paths and the matching size.

We remark that the random process governing the movement of the particles in the network is not a classical random walk over the nodes of the graph. Indeed, if that were the case, a maximum matching would always be reached in polynomial time by a simple cat-and-mouse argument. Instead, a random move of a particle depends on the current matching, which in turn changes when the particle moves. This modest difference can lead to an exponential time gap between convergence to an approximate matching and convergence to a maximum matching. Indeed, exploiting the dependence of the particles' movements on the current matching, we show that there is a family of graphs for which the Prudence algorithm takes exponentially many rounds with high probability to reach a maximum matching starting from a set of configurations that cover almost all possible cases. This family of “bad” graphs is defined as follows (see also Figure 7).

Definition 1 (Bad graph G_n) The bipartite graph has nodes and , and its edges are , for all and , and for all and .

Note that the set of “horizontal” edges , for is the unique perfect matching for .

Theorem 3 The Prudence algorithm requires many rounds with high probability to reach the perfect matching when starting from any -matching in which the two unmatched nodes are in opposite sides of .

The main idea of the proof is to track the positions of the unmatched nodes throughout the course of the algorithm and to lower bound the number of rounds needed before they meet as an adjacent pair.

We first prove a one-to-one correspondence between the Markov process of the state evolution between matchings and a classical random walk on a tree (represented in Figure S1) whose size is exponential in . We show that this classical random walk takes exponential time to reach the root of the tree starting at any one of its nodes, thus providing a lower bound on the convergence time of the algorithm.

We say that a matching of of size is bad if the algorithm requires exponentially many rounds with high probability to converge to the perfect matching when starting from . Observe that all matchings considered by Theorem 3 are bad. The following theorem states that almost all matchings of size are bad.

Theorem 4 The ratio between the number of “bad” matchings and the number of all -matchings of is .

Theorems 3 and 4 show that the algorithm requires exponentially many rounds to converge to the perfect matching of when starting from a set of configurations (the bad matchings) constituting almost all possible cases (the matchings of size ).

Figure 3 compares the performance of the human subjects (red) with that of simulations (blue) on a set of 16-node networks (8 bipartite networks and 8 non-bipartite networks) with unique perfect matchings. The networks are sorted by increasing average completion time, and as a result bipartite networks are labeled from 1 to 8, while non-bipartite networks are labeled from 9 to 16. Each of these networks was tested at least 6 times over all sessions. The vertical axis represents the time (in seconds), and the numerical values of the convergence time of the algorithm are scaled by a constant factor to best match the experimental data.

In an additional experimental session, we tested twenty four subjects connected over small-world, preferential attachment and ring networks as well as over the “bad” graph . The games on the bad graph were never solved, consistent with the prediction of exponentially slow convergence. Furthermore, we found that preferential attachment networks were more difficult to solve than small-world networks (one-tailed Welch's t-test, -value ). Figure 4 shows the affinity between humans' (red) and algorithm's (blue) performance, on this set of 24-node networks: small-world networks (triangles), ring network (diamonds), preferential attachment networks (circles). The -axis shows the indices of the networks sorted by increasing average time to find the perfect matching, and the -axis shows the average time.

Figure 5 shows, by simulation, that the algorithm scales linearly in the size of the network in the case of small-world networks [13], while it scales polynomially for preferential attachment networks [11], [12], and exponentially on the “bad” graph . These results closely resemble the experimental data of the coloring games performed by Kearns et al. [6], where preferential attachment networks resulted in the worst performance among all tested networks, while small-worlds networks appeared to be much easier to solve.

Figure 6 shows the performance of the experimental subjects on networks of nodes, each admitting a perfect matching. In particular, it reports results for single games, and it compares the time to reach a perfect matching of size (red), an approximate matching of size (a -approximate matching, in blue) and a matching of size (a -approximate matching, in green) in each game. The -axis shows the indexes of the games sorted by increasing solution time, while the -axis shows time in seconds. The plot shows (consistent with the theoretical analysis) that a -approximate matching is reached almost immediately in all games, an almost maximum matching is reached quickly, while reaching a perfect matching can take a large amount of time.