Strategyproof Facility Location Mechanisms with Richer Action Spaces

02/18/2020 ∙ by Xiang Yan, et al. ∙ Harvard University 0

We study facility location problems where agents control multiple locations and when reporting their locations can choose to hide some locations (hiding), report some locations more than once (replication) and lie about their locations (manipulation). We fully characterize all facility location mechanisms that are anonymous, efficient, and strategyproof with respect to the richer strategic behavior for this setting. We also provide a characterization with respect to manipulation only. This is the first, to the best of our knowledge, characterization result for the strategyproof facility location mechanisms where each agent controls multiple locations.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

In a classic facility location problem, a social planner chooses to build a facility based on reported locations of agents on a real line. Each agent has one private location and prefers the facility to be built as close to her location as possible. Agents may choose to lie about their locations to influence where the facility is built. It is well-known that choosing the median of reported locations is not only strategyproof but also socially optimal, resulting in the smallest total distance between the facility and agents’ locations. Moulin (1980)’s seminal work fully characterizes all strategyproof facility location mechanisms for this setting.

In many scenarios, for example when each agent represents a community, agents may control more than one locations. The social planner still hopes to select a location to build the facility based on agents’ reported locations. The facility location problem where each agent controls multiple locations was first introduced by Procaccia and Tennenholtz (2009). Choosing the social optimal solution, the median of all reported locations, is no longer strategyproof. Procaccia and Tennenholtz (2009) provided an intuitive, strategyproof mechanism that relabels each reported location by the corresponding agent’s most preferred location and then chooses the median of the relabelled locations. The characterization of all strategyproof mechanisms remains an open question.

Moreover, strategyproof mechanisms so far only guard against one type of strategic behavior, that is agents may lie about their locations in reporting (which we call manipulation in this paper). But when agents control more than one locations, they can choose to hide some locations (hiding) or report some locations more than once (replication) to influence the facility location to their benefit. Strategyproof mechanisms need to be robust against these richer strategic actions. The above-mentioned mechanism provided by Procaccia and Tennenholtz (2009) is not strategyproof w.r.t. replication. Hossain et al. provided examples showing that facility location mechanisms where each agent controls multiple locations may be strategyproof w.r.t. manipulation but not hiding, or vice versa.

In this paper, we fully characterize strategyproof mechanisms w.r.t. the richer strategic actions for facility location problems where each agent controls multiple locations. Intuitively, for each agent reporting multiple locations, one may treat her as an agent controlling a single location, her most preferred location, which is the median of her reported locations. By doing so, all the strategyproof mechanisms for single location agents should also be strategyproof for multiple location settings. One natural guess is that these are all strategyproof mechanisms we desired. This is however not true. We show that there exists other strategyproof mechanisms that depend on not only agents’ most preferred locations but also their other reported locations.

To fully characterize all strategyproof facility location mechanisms where each agent controls multiple locations, we first show that if one cannot distinguish which locations are reported by the same agent, referred as settings with non-identifying locations, all strategyproof mechanisms outputs a constant location that is independent of the reported locations. Then for settings with identifying locations, we show a necessary property for any strategyproof mechanism: for each agent, the mechanism has at most two possible outputs fixing other agents’ reports and the agent’s most preferred location(s). Further adding the Pareto efficiency condition, we derive a full characterization for strategyproof mechanisms. Finally, we compare the result with the characterization for strategyproof mechanisms w.r.t. manipulation only, and discuss the group strategyproofness of the mechanisms.

1.1 Related Work

Dekel et al. (2010)

investigated the framework of mechanism design problems in general learning-theoretic settings. The facility location problem where each agent controls multiple locations is its special case for one-dimensional linear regression, and was first introduced by

Procaccia and Tennenholtz (2009). Some deterministic and randomize mechanisms are provided by Procaccia and Tennenholtz (2009), as well as Lu et al. (2009), and they focused on the approximation ratio of these mechanisms, i.e. the total distance between the facility and all agents’ locations compared with the optimal solution. Hossain et al. extended these studies for considering another strategic behavior, hiding, and provided a strategyproof mechanism w.r.t. both manipulation and hiding. On the other hand, Conitzer (2008); Todo et al. (2011) studied the strategyproofness w.r.t. false-name report, i.e. agents are able to create multiple anonymous identifiers. It is similar to our replication when the identifier of the agent reporting each locations is unknown.

The facility location problems is also a framework for studying “single-peak” preference. As agents with such kind of preference are commonly seen in political decision making like voting (Moulin, 1980), location on networks (Schummer and Vohra, 2002), and resource allocation (Guo and Conitzer, 2010). An extension for the preference structure is the “single-plateau” preference, where the most preferred locations of agents become intervals. Moulin (1984); Berga (1998) provided corresponding characterizations of strategyproof mechanisms for more general social choice settings. Following them, the characterizations were extended to high dimensional Euclidean space (Border and Jordan, 1983; Barberà et al., 1993, 1998) and convex spaces (Tang et al., 2018). These works considered manipulation as the only strategic behavior, and agents are assumed to only report their “peak” or “plateau” preferences. In comparison, our work focuses on facility location settings where agents have “single-plateau” preferences and richer actions spaces.

2 Our Model

Following a brief explanation of notations, we introduce the facility location problem where each agent controls multiple locations, our richer strategic considerations, and desired properties of mechanisms for this problem.

Notations.

Let be the set of first natural numbers. Given real numbers and , let be their median. When

is odd,

, and when is even, , which is a number if and an interval otherwise. Given a set of real numbers, possibly with some identical elements, denotes the size of , and is the median of all numbers in .

Facility Location Problems.

There is a set of agents, . Each agent controls a set of locations, , , on the real line. Each agent is asked to report her set of locations to the principal. Let denote agent ’s set of reported locations,with . can differ from in both size and values. We use to denote the set of locations reported by all agents and to represent the total number of reported locations. The principal seeks a mechanism such that is a proper location for building a facility that will be used by all agents. Each agent incurs a loss

(1)

for using a facility located at

. This loss function means that an agent’s loss is minimized when the facility locates within an interval (the median of the agent’s locations) and strictly increases as the facility moves away from the interval on either side. The following proposition formalizes this property.

Proposition 2.1.

Loss function (1) represents “single-plateau” preferences. Let , where , represent the median interval of agent ’s locations. Then

  • , ,

  • satisfying , , and

  • satisfying , .

Strategic considerations.

Agents want to report to minimize their loss. We allow three types of agent strategic behavior:

  • Manipulation. Each agent may report different value for each of her controlled locations. This is the strategic behavior usually considered in the literature.

  • Replication. Each agent may report one or more of her locations for more than once. Note that even if an agent reports her controlled locations without replication, it is possible that some locations appear for more than once.

  • Hiding. Each agent may choose to no report some of her controlled locations.

The combination of these three types of strategic behavior allows agents to report a set of locations with any size and any value.

Desirable properties of mechanisms.

The principle hopes to find a mechanism that discourages strategic behavior of agents. We define three notions of strategyproofness that we’ll consider. The notions of strategyproofness are with respect to one or more of the three types of strategic behavior.

Definition 2.1.

A mechanism is strategyproof w.r.t. some set of strategic behavior if no agent can achieve less loss by deviating from truthfully reporting to a strategic action in the set, regardless of the reports of the other agents.

Definition 2.2.

A mechanism is group strategyproof w.r.t. a set of strategic behavior if no coalition of agents can simultaneously adopt strategic actions in the set such that every agent in the coalition is strictly better off, regardless of the reports of the other agents.

Definition 2.3.

A mechanism is strong group strategyproof w.r.t. a set of strategic behavior if no coalition of agents can simultaneously adopt strategic actions in the set such that no agent in the coalition is strictly worse off and some agent in the coalition is strictly better off, regardless of the reports of the other agents.

Note that a group strategyproof mechanism must be strategyproof and a strong group strategyproof mechanism must be group strategyproof, but not vice versa. We will discuss the corresponding difference in Section 5.

In addition to strategyproofness, two other properties are also desirable for facility location mechanisms.

Definition 2.4.

A facility location mechanism is anonymous if its output is symmetric w.r.t. all agents.

Definition 2.5.

A facility location mechanism is efficient if its output is Pareto optimal to all agents, i.e. there does not exist another location that is strictly better for at least one agent and not worse for all other agents.

When each agent only controls a single location, denoted by for , Moulin (1980) characterizes that all facility location mechanisms that are anonymous and strategyproof w.r.t. manipulation are of the form

(2)

where are constants; all facility location mechanisms that are anonymous, efficient, and strategyproof w.r.t. manipulation take the form

(3)

where are constants.

3 Strategyproof Facility Location Mechanisms for Non-identifying Locations

We first show that when each agent can control multiple locations, being able to identify which locations are reported by the same agent is necessary for developing non-trivial strategyproof mechanisms. We use the term non-identifying locations to represent the case when one cannot tell which agent reports which location, or more formally, for any reported location , one cannot distinguish the agent who has reported the location. We show below that with non-identifying locations, all strategyproof mechanisms must be the trivial constant mechanisms.

Theorem 3.1.

For facility location problems where each agent controls multiple locations, if the reported locations are non-identifying, then any mechanism that is strategyproof w.r.t. manipulation must output a constant location.

Proof.

For convenience, denote the reported data set as . Since the reported locations are non-identifying, it is possible that for each . Then according to Moulin (1980), any strategyproof mechanism has the form

We next prove that for some , which means the mechanism always returns a constant . Suppose otherwise, w.l.o.g, let . Then we can construct an example where an agent could achieve smaller loss through manipulation. Consider and as the real locations controlled by agent and respectively, where and for . Then truthfully reporting results in and agent suffers a loss of . If agent misreports his locations by manipulating to , then the mechanism will output and her loss becomes , which is strictly smaller if . ∎

Corollary 3.1.

For facility location problems where each agent controls multiple locations, if the reported locations are non-identifying, then any mechanism which is strategyproof w.r.t. manipulation, replication, and hiding, must output a constant location.

Thus, in the rest of the paper, we focus on facility location problems with identifying locations. We note that identifying locations do not conflict with anonymity. Anonymity means that a mechanism’s outcome is not affected by relabelling of the agents, while identifying locations only require that one knows which locations are reported by the same agent and the labels of the agents are not important.

4 Strategyproof Facility Location Mechanisms for Identifying Locations

In this section, we characterize mechanisms that are anonymous, efficient and strategyproof w.r.t. all three types of strategic behavior. We show that these mechanisms take the form of Moulin’s characterization (3) where the median of each agent’s reported locations is used as the representative location for the agent. We further show in Section 4.1 that the family of anonymous, efficient and strategyproof mechanisms remain the same even if only manipulation is considered, when agents control the same number of locations.

We develop our results by first characterizing mechanisms that are anonymous and strategyproof w.r.t. all three types of strategic behavior. Lemma 4.1 shows that from agent ’s perspective, fixing other agents’ reports, any anonymous and strategyproof mechanism can have at most two different outputs if the median of the agent’s reported locations, , is unchanged.

Lemma 4.1.

For facility location problems where each agent controls multiple locations, if a mechanism is anonymous and strategyproof w.r.t. manipulation, replication, and hiding, then from any agent ’s perspective, for any and all with , the output of the mechanism is either or

(4)

for some that satisfies either or . The values of and depend on the reports of other agents as well as and .

Proof.

If a mechanism outputs the optimal location based an agent’s reported locations or a constant location, the agent has no incentive to misreport. Thus, we try to characterize strategyproof mechanisms beyond these two trivial types.

For each and fixed, let be the interval that any location inside is optimal for agent ’s real controlled locations. Through manipulation, replication and hiding, agent is able to misreport any possible location sets. Currently we focus on a special group of misreports, that is agent does not change the median value (either one or two value) of her controlled locations but may misreport any others, i.e. all with .

Suppose there are at least three different outputs for all such which do not belong to the interval , w.l.o.g. let , for . If , then when other agents report , agent with real controlled locations will misreport to obtain a smaller loss. Otherwise, , then when other agents report , agent with real controlled locations will misreport to obtain a smaller loss. This is a contradiction, which means there are at most two different outputs for all such which do not belong to the interval , denoted by . Furthermore, if there exists some such that , then for all , . This means for all , either or .

We further consider the case that . Let and . If , then when other agents report , agent with real controlled locations will misreport to obtain smaller loss. Similarly, if , then when other agents report , agent with real controlled locations will misreport to obtain smaller loss. Thus, we know , and Eqn. (4) is straightforward. ∎

The following example shows that a strategyproof mechanism indeed can have two different outputs, i.e. satisfying Eqn. (4) for .

Example 4.1.

For simplicity, suppose is a unique number for each . For given constants , let be the number of agents who prefer to , and be the number of agents who prefer to . Then the following mechanism is strategyproof:

(5)

where .

Notice that the mechanism in this example is not efficient. We formally prove this observation in the next lemma.

Lemma 4.2.

For facility location problems where each agent controls multiple locations, if a mechanism is anonymous, efficient and strategyproof w.r.t. manipulation, replication, and hiding, then for any , , , there do not exist such that , and satisfies Eqn. (4) for all with .

The proof of Lemma 4.2 makes use of Lemma 4.3.

Lemma 4.3.

For facility location problems where each agent controls multiple locations, suppose a mechanism is anonymous, efficient and strategyproof w.r.t. manipulation, replication, and hiding. If there exist some , , , such that , and satisfies Eqn. (4) for all with , denote two index sets and , then .

Proof of Lemma 4.3.

If , an efficient mechanism should always output the median of the reported locations, which means by definition.

For , w.o.l.g., assume there exists some with , , and , such that satisfies Eqn. (4) for all with . This means there exist at least two special location sets and such that and (for example and ). If , that is and , then is not efficient since any location between and is a better output for both agent and . If , that is and , then is not efficient since any location between and is a better output for both agent and . Thus, the result holds for .

For , assume there exist some , , , and , such that satisfies Eqn. (4) for all with . Consider two special location sets satisfying , . In the following proof, we shall derive contradictions for by induction, and the corresponding analysis for is similar.

Step 1:

Suppose .
This means there is only one , such that , and . Denote as the reported location sets by agents other than agent and , then we rewrite the fact that and .

Step 1.1: Moving inside doesn’t change .
Consider agent ’s another possible report . Due to the efficiency, we know . If , then when agent reports and other agents report , agent with real locations will misreport to obtain a smaller loss. If , then when agent reports and other agents report , agent with real locations will misreport to obtain a smaller loss. This means .

Step 1.2: is the only possible output.
Now for fixed , if there exists such that for some with , , then is not efficient since is a better output for some agents including agent and not worse for others. This means for all with , .

Step 1.3: There exists a beneficial misreport.
Specifically, . However, when agent reports and other agents report , agent with real locations will misreport , which leads to by the anonymity, to obtain a smaller loss. This is a contradiction, meaning it is impossible that .

Step 2.

Suppose we have proved it is impossible that for some , we consider the case .
Let . Similarly, denote as the reported location sets by agents other than agent and , and we rewrite the fact that and .

Step 2.1: Moving inside doesn’t change .
Consider agent ’s another possible report . If , it is not efficient since at least is a better output for all agents as well as agent and not worse for other agents. If , then when agent reports and other agents report , agent with real locations will misreport to obtain a smaller loss. If , then when agent reports and other agents report , agent with real locations will misreport to obtain a smaller loss. This means .

Step 2.2: There must be another possible output .
Now for fixed , if for all with , , specifically we have . Then when agent reports and other agents report , agent with real locations will misreport , which leads to by anonymity, to obtain a smaller loss. This is a contradiction, meaning there must exist such that for some with , .

Step 2.3: and .
If , then due to the efficiency, while since it equals either or . This means when agent reports and other agents report , agent with real locations will misreport to obtain smaller loss. Thus, we must have . In other words, when other agents’ reports change to , there still exist , such that and satisfies Eqn. (4) for all satisfying that is an even number and . And for the corresponding new index set (after replacing by ), we have since and since .

Step 2.4: Repeating previous steps leads to .
Finally, if , we can repeatedly replace the report of the agent whose location set has the leftmost median among those agents in in the same way until , and the same analysis still holds. Since such an index set satisfies , after at most times of such replacing, we must have . However, we have already proved that it is impossible for . By induction, we provide contradictions for , and this completes the proof of . ∎

Proof of Lemma 4.2.

Assume there exist some , , , and , such that satisfies Eqn. (4) for all with . According to Lemma 4.3, we know for any , let , either , or . Let , since . Let . Then for any satisfying , it is not efficient since is a better output for all agent while not worse for other agent. This is a contradiction to the efficiency condition. ∎

Lemma 4.2 indicates that any mechanism that is anonymous, efficient, and strategyproof, must only depend on the optimal location for each agent. In other words, agents only need to report their most preferred locations, i.e. for . With Moulin (1980)’s results, we have the complete characterization.

Theorem 4.1.

For facility location problems where each agent controls multiple locations, a mechanism is anonymous, efficient and strategyproof w.r.t. manipulation, replication, and hiding, if and only if there exist , , for ,

(6)

where with , for .

Notice that if is an interval for some agent , any value in the interval can be regarded as agent ’s optimal location. For simplicity, we only include a tie-breaking rule based on an arbitrary constant , which is independent of () and guarantees the strategyproofness. Moulin (1984) provided more general characterizations for strategyproof social choice mechanisms where each agent reports an interval as her “single-plateau” preference, which deal with the tie-breaking rules carefully.

4.1 Strategyproofness w.r.t. Manipulation Only

In most previous studies on facility location problems where each agent controls multiple locations, manipulation is considered as the only strategic behavior that agents may take. Although some strategyproof mechanisms w.r.t. manipulation are discussed, there is no characterization result. To characterize such strategyproofness, we further assume that each agent control the same number of locations.

Theorem 4.2.

For facility location problems where each agent controls same number of multiple locations, a mechanism is anonymous, efficient and strategyproof w.r.t. manipulation, if and only if there exist , , for ,

where with , for .

The proof of Theorem 4.2 follows in a similar spirit as that of Theorem 4.1, which can be found in Appendix A. Lemma 4.1 still holds but the proof of Lemma 4.2 needs modification because we cannot directly apply Lemma 4.3.

5 Group Strategyproof Facility Location Mechanisms for Identifying Locations

If a mechanism is group strategyproof w.r.t. some set of strategic behavior, then it must be strategyproof w.r.t. the set of strategic behaviors. This means facility location mechanisms which are anonymous, efficient and group strategyproof should satisfy Eqn. (6). We can further show that these strategyproof mechanisms are indeed group strategyproof.

Theorem 5.1.

For facility location problems where each agent controls multiple locations, a mechanism is anonymous, efficient and group strategyproof w.r.t. manipulation, replication, and hiding, if and only if there exist , , for ,

where with , for .

Proof.

Let be any reported locations controlled by agent with for . For any agent and any coalition of agents including , denote their real locations by , and the reported locations of other agents by . W.l.o.g. assume