Abstract

In Braess paradox the addiction of an extra resource creates a social dilemma in which the individual rationality leads to collective irrationality. In the literature, the dynamics has been analyzed when considering impulsive commuters, i.e., those who switch choice regardless of the actual difference between costs. We analyze a dynamical version of the paradox with nonimpulsive commuters, who change road proportionally to the cost difference. When only two roads are available, we provide a rigorous proof of the existence of a unique fixed point showing that it is globally attracting even if locally unstable. When a new road is added the system becomes discontinuous and two-dimensional. We prove that still a unique fixed point exists, and its global attractivity is numerically evidenced, also when the fixed point is locally unstable. Our analysis adds a new insight in the understanding of dynamics in social dilemma.

1. Introduction

Assume that two different points of a network—an origin and a destination—are connected by two possible roads only. The Braess paradox states that, under specific conditions, adding a third road to the network decreases the efficiency of the network. This phenomenon is known in the transportation field and more general scientific literature (see [1–10]). Braess’ paradox occurs because commuters try to minimize their own travel time ignoring the effect of their decisions on other commuters on the network. As a result, the total travel time may increase following an expansion of the network; in fact, even if some commuters are better off using the new link, they contribute to increase the congestion for other commuters.

The theoretical literature of the Braess paradox is particularly productive, especially in transportation, communication, and computer science (far from being exhaustive, see, e.g., [10, 11]). Almost all the existing works have considered the basic network similar to the one presented in this paper, with the addition of a single link. Notably, [12] proposes a broader class of Braess graphs. A measure of the robustness of the dynamic network considering the influence of the flow on other links when certain component (node or link) is removed can be found in [13]. The empirical literature provides evidence in support of the paradox. For example, in [14] examples are reported that occurred on a modeled network of the city of Winnipeg, while [15] focus on a portion of the Boston road network. The experimental works are interested in studying the occurrence of the paradox in a controlled setting (see, e.g., [16–20]), not only in basic but also in augmented networks. This literature provides evidence in strong support of the paradox in some cases (see [18]), while statistically significant, but weaker support in some other cases (see [16, 17, 19, 20]). A comparison of public versus private monitoring using the same participants is performed in [21] to investigate how the type of monitoring affects route choice. Interest in possible behaviors and composition of the population facing the basic network can be found in [22] which analyze the data gathered from the observation of an experiment with human participants, codes artificial behaviors emerged by mean of grounded theory, and uses ABM simulations. For a review on ABM with special focus on spacial interactions and networks see [23].

The paradox has also been presented in a dynamic game framework. A discrete time dynamic population model with social externalities and two available choices is studied in [24] simulating an adaptive adjustment process, and in particular with impulsive commuters in [25].

Following the clinical psychology literature (see [26]), impulsive commuters have been introduced in the analysis of binary choices with externalities in [25, 27] when considering commuters whose switching rate only depends on the sign of the difference between payoffs, no matter how much they differ. This approach has been used to consider choices both in small and in large groups in [28] and also when considering the introduction of a third choice which dramatically changes the dynamics in terms of complexity (as analyzed in [29]).

Analyzing the Braess paradox in terms of a ternary choice game shows new interesting dynamical characteristics that are investigated in [30], with a particular interest in the coexistence of several equilibria. Nevertheless, as illustrated in [22, 31], considering only impulsive commuters is not sufficient to describe the dynamics observed in experiments with human participants. Furthermore, as impulsive commuters consider only the sign and not the size of difference between costs, the population dynamics does not depend on the cost functions as long as the indifference point remains the same. In this paper, we consider a different behavior suggested by Amnon Rapoport (We are grateful to him for this helpful suggestion.) and which was used in [22] for artificial commuters. Although the full analysis of a homogenous—yet different—population seems to provide a limited contribution, it may be a step forward to analyze the aggregate decision behavior in social dilemmas. In fact, in order to analyze heterogeneous populations, it is important to well understand the dynamics properties of component behaviors. In particular, we consider commuters which are concerned not only about the sign difference in payoffs but also on the relative difference. This kind of behavior is similar to the one considered in [32], where the propensity to switch choice is modulated by the difference between payoffs. Unlike the impulsive adjustment process, which makes the commuters change their choice as soon as a better choice occurs regardless of the difference, we consider a decision strategy which prescribes a change to the best available choice proportionally to the reported difference. Intuitively, this strategy seems to be more robust. The main result of the present work is that it is in fact robust. By contrast to the strategy considered in [30], which may lead to a stable cycle of any period, the strategy considered in this paper leads to a globally attracting fixed point. Depending on the parameters, this unique fixed point may be locally stable or unstable. However, the trajectories are convergent, in a few steps, to the fixed point. This result is rigorously proved for the case of a binary choice problem. For the ternary choice problem an analogous result is evidenced as well.

The plan of the work is as follows. In Section 2 the formalization of the dynamic model is reported. Section 2.1 includes the detailed description of the case with only two roads, which is represented by a one-dimensional piecewise smooth continuous map. We prove the existence of a unique fixed point, which may be locally stable or unstable in one partition. We prove that in any case it is globally attracting, that is, all the trajectories are ultimately converging to the fixed point. The case extended to three roads is considered in Section 3 and it is reduced to a two-dimensional piecewise smooth discontinuous map. We prove the existence of a unique fixed point, which may be locally attracting or not. We have only numerical evidence that also in this more general case, the fixed point is globally attracting. This can be rigorously proved for the particular cases in which the fixed point belongs to the boundaries of the domain of interest. The last section is devoted to the conclusions.

2. The Dynamic Model with Two Roads

The Braess paradox can be illustrated by Figure 1 as follows. Assume there is a unitary mass of commuters from start () to end () and there are two roads: one passing through left () and the other through right (). The cost of each road is given by the sum of the time spent along each segment. In this network, segments - and - do not depend on traffic and they cost and , respectively. On the contrary, time spent along segments - and - is proportional to the number of travellers and therefore, they cost and , where is the fraction of the population using segment -. At the Nash equilibrium commuters are distributed in such a way that both roads (-- and --) have equal cos;, that is, a fraction of commuters will choose -- and a fraction will choose --. For example, assume as in [22] that and . Then, the symmetry of costs makes the population to split exactly into two equal fractions at the Nash equilibrium (), so that the travel time for both roads is. Now we assume that a very fast road is built connecting to at cost . As we will see in Section 3, at the new Nash equilibrium commuters are distributed in such a way that the three roads (--, ---, and --) have equal cost; that is, a fraction of commuters will choose --, a fraction will choose --, and the rest will choose ---. In the numerical example, this means that the entire population will choose --- (since in the other two roads the fraction of commuter is ) with a total travel cost of . According to [19], this is considered paradoxical as it shows how adding extra capacity to a network can reduce overall performance. More correctly, it is just counterintuitive and the mathematical reason is that there is a distinction between Nash equilibria and optima.

(a)

(b)

(a)
(b)

Figure 1 

Basic network with the addition of the link from to in (b) and without in (a).

We consider the original network first made of two roads only (i.e., without the link -) as in Figure 1(a).

As in [24] this process can be modeled as a discrete time dynamical system considering commuters facing a binary choice between paths -- and --. For the sake of brevity these paths will be denoted as and , respectively. To this purpose, assume the set of commuters is normalized to the interval . At each time period , indicates the fraction of commuters choosing and the fraction of those choosing , given by . Therefore, the travel times are given by where and can be written as This way, the travel times on the two roads are equal (and the commuters are indifferent) when ; that is, when , where We say that the threshold is feasible when . However, (resp., ) clearly corresponds to the case in which the whole population chooses action (resp., ), and we shall see that this is indeed satisfied in the model described in the next subsection. We obtain the following constraints on the parameters in order to have , which we assume henceforth

2.1. The One-Dimensional PWS Map and Its Dynamics

Commuters are homogeneous and at each period decide their strategy considering the previous period travel costs assuming they will remain the same. At time the fraction is common knowledge and each commuter observes the respective travel costs either or , depending on their choice. The commuters decide their future action at time comparing the costs and . We assume the commuters are nonimpulsive and myopically minimize their costs. That is, they change strategy proportionally to the difference between their chosen road and the best one, where the best road is the fastest. We introduce parameters and as switching rates, that is, the propensity for moving either to or to , respectively. With nonimpulsive commuters these rates are adjusted by the normalized difference of travel costs. The difference is normalized by the largest value of this difference, denoted by and obtained in correspondence of . Therefore we have as already introduced in (4). Thus commuters are switching road proportionally to the side of the payoff difference. Similarly we can give a reason with respect to the difference to get a normalizing value. Now the largest value is obtained in correspondence of ; that is, , and we get as already introduced in (4).

The resulting dynamics is with and the map is defined as Substituting from (2) and rearranging we obtain From the definition of given in (3), it can also be written as at which we have , it is also and thus is continuous in and it is a fixed point as . We have the following proposition.

Proposition 1. Mapis continuous in and under the conditions given in (4) is the unique fixed point.

Proof. Only the uniqueness is left to prove. The equation has two solutions: and which is larger than and thus a virtual fixed point. The equation has two solutions: and which is smaller than and thus a virtual fixed point. Thus, the unique solution is .

Clearly the fixed point may be locally stable or instable, depending on the slopes of the functions on the right and left sides of . However, we shall see that in any case it is globally attracting. That is, any i.c. in the interval has a trajectory which converges to .

Proposition 2. The fixed point is globally attracting.

Proof. Even if is not smooth in its fixed point (as indeed is a kink point of this map), the two components and are smooth in so that the left and right side derivatives of in are well defined: and .
We have while From we have that on the left side of the fixed point the function is convex, while on the right side of the fixed point the function is concave.
Notice that only depends on , and it is positive under our assumptions. Depending on the values of the ratio we can classify the right/left stability of the fixed point . From (9) we clearly have always ; thus the stability/instability on the left side of the fixed point depends on the condition , which holds if and only if Similarly, from (11) we clearly have always , and the stability/instability on the right side of the fixed point depends on the condition , which holds if and only if And clearly we have the following sequence of inequalities (to be used below): To investigate the local stability of the fixed point we consider three intervals for :(i)(ii)(iii).
In case (i) it is , moreover, it is easy to see that for any it is and thus the function on the right side of the fixed point is strictly decreasing and expansive: each point on the right side of the fixed point is mapped to the left side of the fixed point in one iteration. While on the left side of the fixed point the convex function has (as this holds for which is satisfied in this interval). This implies that the point of local minimum on the left side, say , is smaller than the fixed point, . Then the fixed point is unique and the map is a contraction in the interval , and an absorbing interval which is mapped into in one iteration is given by when , or when . Any initial condition belonging to has a trajectory which is mapped into in a finite number of iterations. Thus is globally attracting. An example is shown in Figure 2(a).
Similarly we can reason when belongs to the third interval. In fact, in case (iii) it is , moreover, it is easy to see that for any it is and thus the function on the left side of the fixed point is strictly decreasing and expansive: each point on the left side of the fixed point is mapped to the right side of the fixed point in one iteration. While on the right side of the fixed point the concave function has (as this holds for which is satisfied in this interval). This implies that the point of local maximum on the right side, say , is larger than the fixed point, . Then the map is a contraction in the interval , and an absorbing interval which is mapped into in one iteration is given by when , or when . Any initial condition belonging to has a trajectory which is mapped into in a finite number of iterations. Thus is globally attracting. An example is shown in Figure 2(b).
When belongs to the interval in (ii), , we have that the fixed point is locally stable on both sides. Moreover, inside this interval we can distinguish three regions, as already showed in (15), that is:(a) with and , as in Figure 3(a)(b) with and , as in Figure 3(b)(c) with and , as in Figure 3(c).
In case (a) we have that the point of local minimum on the left side, say , is smaller than the fixed point, . Then the map is a contraction in the interval , and on the right side the map is decreasing, so that an absorbing interval which is mapped into in one iteration is given by when , or when . Any initial condition belonging to has a trajectory which is mapped into in a finite number of iterations. Thus is globally attracting.
In case (c) we can reason similarly, we have that the point of local maximum on the right side, say , is larger than the fixed point, . Then the map is a contraction in the interval , and on the left side the map is decreasing, so that an absorbing interval which is mapped into in one iteration is given by when , or when . Any initial condition belonging to has a trajectory which is mapped into in a finite number of iterations. Thus is globally attracting.
In case (b) we have that the point of local minimum on the left side, say , is smaller than the fixed point, and also that the point of local maximum on the right side, say , is larger than the fixed point, . Then the map is a contraction in the interval which includes the fixed point, and an absorbing interval which is mapped into in one iteration is given by , or or or , depending under obvious conditions on the values of and . When then any initial condition belonging to has a trajectory which is mapped into in a finite number of iterations. Thus is globally attracting.

(a)

(b)

(a)
(b)

Figure 2 

Qualitative shapes of the map. In (a) case (i), when , the fixed point is locally unstable on the right side. In (b) case (iii), when , the fixed point is locally unstable on the left side.

(a)

(b)

(c)

(a)
(b)
(c)

Figure 3 

Qualitative shapes of the map in case (ii), when and the fixed point is locally stable on both sides.

We notice that in Proposition 2 we prove that when the fixed pointis internal to the interval then it is globally attracting. Let us turn to comment the two extrema, which occur when or . These solutions clearly exist and are also globally attracting. In fact, the case corresponds to and only the function is applied in the interval and from the concavity of the function we have that this fixed point is globally attracting even if the eigenvalue is equal to 1. The other case corresponds to , and only the function is applied in the interval and from the convexity of the function we have that this fixed point is globally attracting even if the eigenvalue is equal to 1.

3. The Dynamics in the Expanded Network with Three Roads

As it is well known, the Braess paradox introduces a new link and shows how this opportunity worsens both the individual and collective payoffs. This link introduces a new alternative: now the commuters are facing three choices, where the new one consists of path --- (see Figure 1(b)), which will be denoted as in the following. Models with three choices were introduced in [29] as it concerns discrete time and [33] as it concerns continuous time. In this paper we analyze the dynamics when the travel time for the new resulting link varies in a range that makes this link a dominant choice for some values and a dominated choice for others. It is worth to observe that in particular when the travel time on link - is zero, the paradoxical result still holds.

Now assume that commuters switch road whenever travel time has become larger; that is, the other road is less costly and therefore more attractive. We indicate by the fraction of population choosing action , by the fraction choosing action , and by the fraction choosing action , with the constraint . The travel times are given, respectively, as By using the constraint let us define the fraction of population choosing as , so to consider only two variables, , and the phase space is the triangle After substituting the expression of in the travel times we get The vertices of the phase space , , and represent states in which the whole population chooses, respectively, actions , , and .

Since commuters are interested in minimizing travel time, we denote by the region in which the choice (where is , , or ) is preferable. These regions are called dominance regions of . We have the following definition of the dominance regions: By using the expressions in (18) we get the following definitions for the regions where we have introduced The boundaries of the three regions are given by segments of straight lines of equations and it is immediate to see that the point satisfies also the third equation; that is, it belongs to all the three straight lines, for .

Using the geometric representation in the triangle the dominance regions are illustrated in Figure 4 (via an example) using different colors. Three cases are shown, for different values of the parameter , in order to illustrate how the region depends on the parameters of the third branch.

(a)

(b)

(c)

(a)
(b)
(c)

Figure 4 

Dominance regions in the three choices network with different values of the cost parameterof the map: at , , , and (a) , (b) , (c) .

The point is the unique point on the boundaries of all the three regions. Thus for our model we have to require , , and , leading to . In order to have three regions in it must be , , and . The equality in one of these three conditions implies the disappearance of one region. For example, when , then and , leaving only the two regions and and so on.

The conditions , , and lead to the constraints that will be assumed satisfied henceforth: We notice that from the third condition in (23) we have the values of for which the third region is really present in ; that is: Commuters are homogeneous and minimize their next period travel cost. At time the vector becomes common knowledge, and each commuter can observe the travel costs , , and . We assume that if at time a fraction chooses path , a fraction chooses action and travel costs are such that and , then both a fraction of the commuters who chose action and a fraction of the () commuters who chose action will switch to path in the next time period . The fraction of switching commuters is given by the switching propensities , , modulated by the relative differences in payoffs. This happens whenever a path gives a smaller travel cost. In other words, at any time all the commuters decide their future action at time comparing the costs , , and . As in the one-dimensional case, here also the differences of travel costs are normalized dividing each difference by the constants giving the related maximum value, given in the different cases as follows: The resulting dynamics are described by a two-dimensional map defined as where By construction we have that maps into itself. In fact, to show that it is enough to show that given then . From we have that and . Then if , we have , and ; thus . If we have and ; thus . If we have , and ; thus .