I Introduction
Noncooperative game theory represents a contemporary and pervasive paradigm for the modelling and optimization of modern multiagent systems, where agents are typically modelled as rational decisionmakers that interact and selfishly compete for shared resources in a stationary environment. Here, the (generalized) Nash equilibrium solution concept [facchinei2007generalized] denotes a desired outcome of the game, which is typically selflearned by the agents through iterative procedures alternating distributed computation and communication steps [salehisadaghiani2016distributed, salehisadaghiani2019distributed, ye2017distributed, gadjov2021exact].
Realworld scenarios, however, are rarely stationary. This fact, along with recent developments in machine learning and online optimization [jadbabaie2015online, Bogunovic2016, shahrampour2017distributed, davis2019stochastic, dixit2019online, simonetto2019personalized, dall2020optimization], has fostered the implementation of online multiagent learning procedures, thus contributing in growing the interest for games where the population of agents ambitiously aim at tracking possibly timevarying Nash equilibria online. A recent research direction, indeed, established the convergence of online distributed mirror descenttype algorithms in strictly monotone GNEP [tampubolon2020coordinated]
, aggregative games with estimated information
[tampubolon2019convergence], or in pricebased congestion control methods for generic noncooperative games [tampubolon2020robust]. Conversely, [mertikopoulos2019learning] focused on the prediction of the longterm outcome of a monotone NEP, also extended to the case of delays in the communication protocol [zhou2018multi]. The convergence of noregret learning policies with exponential weights in potential games within a (semi)bandit framework was explored in [cohen2017learning], while [cardoso2019competing] introduced an algorithm with sublinear Nash equilibrium regret under bandit feedback for timevarying matrix games, and [duvocelle2018learning] showed that, in case of a slowlyvarying monotone NEP, the dynamic regret minimization allows the agents to track the sequence of equilibria.Unlike the aforementioned literature, in this paper we focus on both static and timevarying, nonmonotone generalized Nash equilibrium problems that exhibit symmetric interactions among the agents. Prominent examples can be found in smart grids and demandside management [kattuman2004allocating, zhu2011multi, cenedese2019charging] in case one considers a fair electricity market; in congestion games [rosenthal1973class, altman2007evolutionary], in which shared resources incur costs that depend on the number of users occupying them; or in coordination control problems [zhang2010cooperative, fabiani2018distributed], where barrier functions typically enforce distancebased constraints. Such a symmetric structure brings numerous advantages to the GNEP that enjoys it. Among them, the underlying GNEP is known to be potential with associated potential function [ui2000shapley, la2016potential], which implicitly entails the existence of a GNE. Nonetheless, such a potential function frequently enables for the design of equilibrium seeking algorithms with convergence guarantees (especially in nonconvex setting [heikkinen2006potential, fabiani2019multi, cenedese2019charging]).
However, unless one has a deep knowledge on the main quantities characterizing the symmetric interactions of the GNEP at hand or, conversely, the agents’ cost functions are designed to result in a potential game [li2013designing], finding the formal expression of the potential function is known to be a hard task [la2016potential, Ch. 2][hobbs2007nash]. Moreover, in some cases it may simply be either unavailable or unknown. In many innetwork operations that require some degree of coordination, indeed, it is highly desirable that the parameters of the agents’ cost function, which reflect local sensitive data, stay private. Specifically, our work is motivated by the following application, thoroughly discussed in §VI as case study for numerical simulations.
Ia Motivating example: Ridehailing with mobility as a service orchestration
With the growing business related to ridehailing, a maas coordination platform appears indispensable to contrast the traffic congestion due to the increasing number of vehicles dispatched on the road, while at the same time facilitating the competition among service providers [pandey2019needs, diao2021impacts]. Specifically, let us consider a scenario with ridehailing type firms, such as Lyft, Uber, Didi Chuxing, Via, or Juno, which compete to put the most vehicles (a capped local resource, ) on the road to attract the most customers. During the day, each company aims at maximizing its profit , which is implicitly related to how many cars it could currently put on the road to meet customer needs, properly scaled and discounted to account for, e.g., refusals rates or the time of the day, . To this end, bigger companies can naïvely be induced to dispatch as many cars as they own. However, this may cause traffic congestion, thus reducing the quality of the service provided, and therefore lessen what the company can charge for each ride. In fact, by leveraging their own experience, those big companies may estimate how many cars actually get customers on top of the available as, e.g., a concave function , with tuned accordingly. Therefore, assuming the same fare applies per average trip to each costumer, the profit function of the th lead company can read as On the other hand, the strategies of smaller companies are typically less affected by traffic congestions, since the quality service is generally worse in the sense that they can dispatch a little number of cars on the road. In this case, their direct experience may suggest that the number of cars that actually get customers on the available can be modelled as a convex function , thus reflecting the fact that the larger the number of deployed cars, the larger the possibility to cover enough space to be attractive. The overall profit hence reads as In addition to the profit, however, the companies also incur in costs that have to be minimized and vary during the day, such as gas consumed or miles travelled. By assuming, for instance, the same cost associated to each vehicle per average trip, the overall cost can be modelled as a linear function of the number of cars such as Finally, for competition purposes, companies and with the same size would like to roughly offer the same service, so that needs to be minimized as well. At the same time, it may happen that a certain company aims at offering a better service than company , and hence coupling constraints in the form arise, while the ridehailing firms as a whole can put on the road a maximum of cars, so that , for some minimum service lower bound , with .
In the proposed scenario where the firms exhibit symmetries in the mixed convexconcave cost functions, the maas platform aims at coordinating the whole ridehailing service while avoiding traffic congestion. This can be achieved, for instance, by imposing extra fees, incentives or restrictions to the companies, possibly according to their size and turnover. However, note that the parameters and affecting the cost function of each firm, which are hence key to drive its strategy, can not be disclosed to the maas platform, since they represent sensitive information, as opposed to the incurred cost that can be estimated directly, as it depends on mileage, fuel consumption and number of deployed cars, all data that are somehow publicly available (see, e.g., [nyc_ridehailing]). Therefore, a possible strategy for the maas platform establishes to learn those timevarying parameters by leveraging feedback collected from users, e.g., on the price they are charged , and then design tailored incentives for the coordination.
IB Main contributions
We design a semidecentralized scheme that allows the agents to compute (or track in a neighbourhood) a GNE of a nonmonotone GNEP that admits an unknown potential function, both in static and timevarying setting (§II). Specifically, in the outer loop of the proposed twolayer algorithm we endow a coordinator with an online learning procedure, aiming at iteratively integrating the (possibly noisy and sporadic) agents’ feedback to learn some of their private information, i.e., the pseudogradient mappings associated to the agents’ cost functions. The goal is to drive the population to an GNE. The reconstructed information is thereby exploited by the coordinator to design parametric personalized incentives for the agents [simonetto2019personalized, ospina2020personalized, notarnicola2020distributed]. On its side, the population of agents receives those personalized incentives, computes a solution, i.e., a vGNE, to an extended game by means of available algorithms in the inner loop, and then returns feedback measures to the coordinator.
Unlike the proposed problem setting, we stress that [fabiani2021nash] considered a specific class of stationary nonmonotone GNEP only, i.e., the quadratic one, where a feedback to the coordinator was provided at every iteration. Furthermore, we highlight that GNE seeking in nonmonotone GNEP is a hard task even in a stationary setting, whose underlying literature is not extensive. Examples of solution algorithms, tailored for static NEP or generic VI (hence possibly not amenable to distributed computation), can be found in, e.g., [konnov2006regularization, yin2009nash, yin2011nash, konnov2014penalty, lucidi2020solving]. In addition, our semidecentralized scheme may also fit in a Stackelberg game framework, in which the leader does not control any decision variable, albeit aims at minimizing the unknown potential function on the basis of optimistic conjectures on the followers’ strategies [kulkarni2015existence, fabiani2020local]. Here instead, we leverage the symmetry of interactions characterizing the agents taking part into the nonmonotone GNEP to design parametric personalized incentives, which play a crucial role in the convergence of the algorithm, as they bring a twofold benefit: i) enabling the agents for the computation of a vGNE in the inner loop by acting as a convexification terms for their cost functions; and ii) boosting the convergence and/or lessening the tracking error through a fine tuning of few parameters (§III). As main results, in the static case we show that the proposed algorithm converges to a GNE by exploiting the asymptotic consistency bounds characterizing typical learning procedures for the coordinator, such as ls or gp (§IV). Conversely, in the timevarying setting we show that the fixed point residual, our metric for assessing convergence, asymptotically behaves as , i.e., the proposed semidecentralized scheme allows the agents to track a GNE in a neighbourhood of adjustable size (§V). We corroborate our findings on a numerical instance of the ridehailing service with maas orchestration in §VI. The proofs of theoretical results are all deferred to Appendix A–D.
Notation: , and denote the set of natural, real and nonnegative real numbers, respectively. is the space of
symmetric matrices. For vectors
and , we denote and . With a slight abuse of notation, we also use . is the class of continuously differentiable functions. The mapping is monotone on if for all ; strongly monotone if there exists a constant such that for all ; hypomonotone if there exists a constant such that for all . If is differentiable, denotes its Jacobian matrix. Throughout the paper, variables with as subscript do not explicitly depend on time, as opposed when is an argument.Ii Problem formulation
In this section, we first formally introduce the multiagent equilibrium problem addressed, and then we discuss some key points that essentially motivate our solution algorithm.
Iia Noncooperative GNEP with symmetric interactions
We consider a noncooperative game , with agents, indexed by the set . Each agent controls its locally constrained variable and, at every discrete time instant , aims at solving the following timevarying optimization problem:
(1) 
for some function , , which denotes the private individual cost, whose value at time can be interpreted as the (dis)satisfaction of the th agent associated to the collective strategy . The collection of optimization problems in (1) amounts to a GNEP, where every is a map stacking coupling, yet locally separable, constraints among the agents. Let us first define the sets and , with , and then let us introduce some standard assumptions.
Standing Assumption 1.
For each , and for all ,

The mapping is of class and has a Lipschitz continuous gradient;

is a nonempty, compact and convex set, is a convex and of class function.
The feasible set of the timevarying GNEP thus coincides with [facchinei2007generalized, §3.2]. In the proposed timevarying context, we are then interested in designing an equilibrium seeking algorithm for the game , according to the following popular definition of GNE.
Definition 1.
(Generalized Nash equilibrium [facchinei2007generalized]) For all , is a GNE of the game if, for all ,
(2) 
A collective vector of strategies is therefore an equilibrium at time if no player can decrease their objective function by changing unilaterally to any other feasible point. For the remainder, we make the following assumption on the pseudogradient mapping , which is formally defined as .
Standing Assumption 2.
For every and , .
Roughly speaking, Standing Assumption 2 establishes that each pair of agents influences each other in an equivalent way. For the mapping , this entails the existence of a differentiable, yet possibly unknown, function such that , for all and [facchinei2007finite, Th. 1.3.1], which coincides with a potential function [facchinei2011decomposition] for and can be characterized as stated next.
Lemma 1.
For all , is Lipschitz continuous, while is weakly convex, with .
Note that is a smooth function, which in principle may be nonconvex. Let be the set of its (local and global) constrained minimizers, assumed to be nonempty, and be the set of its constrained stationary points, with , for all . We stress that the nonemptiness of guarantees the existence of a GNE for , since any satisfies the relation in (2).
IiB Main challenges and technical considerations
In the considered formulation, we identify three main critical issues, both technical and practical, that rule out the possibility to compute a GNE for the timevarying GNEP in (1) through standard arguments, thus fully supporting the need for a tailored learning procedure as the one introduced later in the paper.
First, the timevarying nature of the optimization problems in (1) calls for an answer to the thorny question on whether there exist online learning policies that allow agents to track a Nash equilibrium over time (or to converge to one if the stage games stabilize). Even in the case of a potential game with known potential function, this is a challenging problem [cohen2017learning].
In addition, despite the symmetry of interactions among agents, we note that for all the mapping may not be monotone, a key technical requirement for the most common solution algorithms for GNEP available in the literature, which compute a GNE by relying on the (at least) monotonicity of the pseudogradient mapping [salehisadaghiani2016distributed, salehisadaghiani2019distributed, ye2017distributed, gadjov2021exact].
Finally, we stress that Standing Assumption 2, albeit quite mild and practically satisfied in several realworld scenarios [rosenthal1973class, kattuman2004allocating, zhu2011multi, zhang2010cooperative], is key to claim that the underlying GNEP is potential, a fact that typically helps in designing Nash equilibrium seeking algorithms with convergence guarantees (especially in nonconvex/nonmonotone setting, e.g., [heikkinen2006potential, fabiani2019multi, cenedese2019charging, lei2020asynchronous])). However, unless one has a deep knowledge of the GNEP at hand, finding the formal expression of the potential function is known to be a hard task [la2016potential, Ch. 2]. Thus, we assume to do not have an expression for that can be exploited directly for the equilibrium seeking algorithm design.
To address these crucial issues, we design personalized feedback functionals in the spirit of [simonetto2019personalized, notarnicola2020distributed, ospina2020personalized, fabiani2021nash], which are used as “control actions” in the twolayer semidecentralized scheme depicted in Fig. 1. Specifically, our goal is to steer the noncooperative agents to track minimizers of the unknown, timevarying function , i.e., a GNE of the game , according to Definition 1. Any can indeed be interpreted as a collective strategy that minimizes the (dis)satisfaction of the of agents, measured by the function .
Iii Learning algorithm with personalized incentives
We here describe the main steps of the proposed semidecentralized learning procedure, also discussing how the design of personalized incentive functionals promises to be key in addressing the challenges introduced in the previous section.
Iiia The twolayer algorithm
The proposed approach is summarized in Algorithm 1. Specifically, in the outer loop a central coordinator aims at learning online (i.e., while the algorithm is running) the unknown, timevarying function (or its gradient mapping, ) by leveraging possibly noisy and sporadic agents’ feedback on the private functions ’s (S0). On the basis of the estimated , at item (S1) the coordinator designs personalized incentive functionals , and at item (S2) induces the noncooperative agents to face with an extended version of the GNEP in (1), i.e., , with in place of . Under a suitable choice of the personalized incentives, we will show that they act as regularization terms, as well as they tradeoff convergence and robustness to the inexact knowledge of function and its gradient. Specifically, such incentives enable for the practical computation of an equilibrium of the extended game at item (S2) through available solution algorithms for GNEP [salehisadaghiani2016distributed, ye2017distributed, gadjov2021exact].
Note that standard procedures in literature typically returns a vGNE [cavazzuti2002nash, facchinei2007generalized], which coincides to any solution to the GNEP that is also a solution to the associated VI, i.e., any vector such that, for all ,
(3) 
where the mapping is formally defined as , and . For these reasons, in referring to the computational step (S2), we tacitly assume that the agents compute a vGNE of the extended game .
Finally, at item (S3) the agents return feedback measures and their equilibrium strategies, , with
, for some random variable
, to the central coordinator, thus indicating to what extent the current equilibrium (dis)satisfies the entire population of agents.IiiB Personalized incentives design
In view of Standing Assumption 2 and the consequences it brings, e.g., the fact that , a natural approach to design the personalized incentives seems to iteratively learn and point a descent direction for the unknown function , thus implicitly requiring one to estimate the pseudogradients , at every . Along the line of [ospina2020personalized, fabiani2021nash], we assume the central coordinator being endowed with a learning procedure such that, at every outer iteration (Algorithm 1, item (S0)), it integrates the most recent agents’ feedback to return an estimate of the pseudogradients, . A possible personalized incentive functional can hence be designed as
(4) 
where , for some parameters , , for all . Unlike what one might expect, each requires a positive sign for the gradient step . However, note that this fact is not uncommon – see, e.g., the recent Heavy Anchor method [gadjov2021exact, Eq. (7)]. Moreover, in the next sections we will also discuss how such a choice enables us to boost the convergence of Algorithm 1 (as shown, for example, in [fabiani2021nash, §V] on a numerical instance of a quadratic hypomonotone GNEP) or lessen the tracking error, through a fine tuning of the stepsize .
Thus, once the parametric form in (4) is fixed, we design suitable bounds for and in such a way that the sequence of GNE, , monotonically decreases and converges to some point in . As stressed in the previous section, the gain is crucial to enable for the computation of a vGNE at item (S2) in Algorithm 1, as stated next.
Proposition 1.
Let for all . Then, with the personalized incentives in (4), the mapping is strongly monotone, for all .
Thus, at every , in (S2) the population of agents computes the (unique, see [facchinei2007finite, Th. 2.3.3]) vGNE associated to the extended version of the GNEP in (1), .
In the remainder of the paper, we will consider both the stationary and timevarying case of online perfect and imperfect reconstruction of the pseudogradient mappings, also analyzing the results with different learning strategies . To this end, a key quantity will be the fixed point residual , whose norm “measures” the distance to the points in when the function is fixed in time, as stated next.
Lemma 2.
Let be the sequence of vGNE generated by Algorithm 1 with , assume perfect reconstruction of the mapping , and that for some . Then, is a stationary point for the function , i.e., .
It is fundamental to appropriately choose and to drive the sequence of vGNE along a descent direction for the unknown , and ensuring . In case of imperfect reconstruction of , or in the timevarying setting, we also adopt the average value of over a certain horizon of length , i.e., , , as a metric for the convergence of the sequence generated by Algorithm 1 to the stationary point set.
We remark here that, on the one hand, finding the stationary points is the general goal in nonconvex setting [scutari2017parallel], and on the other hand, since Algorithm 1 generates monononically decreasing values for , the application of simple perturbation techniques (e.g., [escape]) can ensure that the stationary points to which we converge are in practice constrained local minima for , namely points belonging to , and therefore GNE of the GNEP in (1), according to Definition 1.
Remark 1.
The bounds on the parameters and provided in the paper assume the knowledge of the constant of weak convexity of , . However, as long as the coordinator is endowed with a learning policy, one may include this additional condition in the learning process, thus obtaining bounds that depend on , the estimate of .
Iv The stationary case
We start by discussing the case in which each in (1) is fixed in time, thus implying that . First, we analyze the case of perfect reconstruction of the pseudogradient mappings (§IVA), and then we investigate their inexact estimate (§IVB). Here, our result will be of the form in case the reconstruction error is nonvanishing. Otherwise, (§IVC), thus recovering the results shown in §IVA.
Iva Online perfect reconstruction of the pseudogradients
In case the learning procedure enables for , , by adopting the personalized incentives in (4) at every outer iteration , we have the following result.
Lemma 3.
Let and , for all . Then, with the personalized incentives in (4), the vector is a descent direction for , i.e., .
Then, if (resp., ) is large (small) enough, at every iteration of Algorithm 1, the personalized functionals in (4) allow to point a descent direction for the unknown (dis)satisfaction function . Next, we establish the convergence of the sequence of vGNE generated by Algorithm 1.
Proposition 2.
By introducing , from the first step of the proof of Proposition 2 we have that , which points out that a fine tuning of the term allows us to boost the convergence of Algorithm 1 to some point in (also observed on a numerical example in [fabiani2021nash, §V]). This essentially explains the choice for a positive sign in the gradient step of (4). However, due to the presence of noise in the agents’ feedback , it seems unlikely that the online algorithm is able to return a perfect reconstruction of , at least at the beginning of the procedure in Algorithm 1.
IvB Inexact estimate of the pseudogradients
At every outer iteration , we assume the coordinator has available agents’ feedback , , and , to estimate the gradients (and hence the mapping ). The value of reflects situations in which the coordinator gathered information before starting the procedure (), or it obtains sporadic feedback from the agents (). Without restriction, we make the following, standard assumption on the reconstructed mapping directly, rather than on each single gradient [dixit2019online, ospina2020personalized, dall2020optimization].
Assumption 1.
For all and , , and, for any , there exists and available agents’ feedback such that , for some nonincreasing function such that , for all .
With Assumption 1, the reconstruction error on made by
is bounded with high probability
by some function of the available agents’ feedback. Then, after defining the quantities and , we have the following result.Lemma 4.
In case of inexact estimate of the pseudogradients, the vector is not guaranteed to be a descent direction for the unknown function . In fact, the term rules out the possibility that the LHS in (5) is strictly negative, albeit it can be made arbitrarily small through by an appropriate choice of the stepsize . As in §IVA, the following bound characterizes the sequence of vGNE generated by Algorithm 1.
Theorem 1.
Let Assumption 1 hold true for some fixed , and , for all . Moreover, let some be fixed, and, for any global minimizer , . Then, with the personalized incentives in (4), the sequence of vGNE , generated by Algorithm 1, satisfies the following relation with probability
(6) 
Here, , , and is the number of available agents’ feedback at the th outer iteration, .
Roughly speaking, Theorem 1 establishes that, with arbitrarily high probability, the average value of the residual over a certain horizon is bounded by the sum of two terms, which depend on the initial distance from a minimum for the unknown function , and the reconstruction error . Note that the terms in the RHS can be made small by either choosing a small stepsize , in order to make close to zero, or tuning the product close to one, thus leading to a large . This latter choice, however, would increase the term involving the suboptimal constant , thus requiring an accurate tradeoff in tuning the gain and the stepsize . In the stationary case, to foster not exceedingly aggressive personalized actions the coordinator may then want to match the lower bound for , while striking a balance in choosing to possibly boost the convergence of Algorithm 1.
For simplicity, let us now assume that is a constant term. From Assumption 1, , and hence We note that, as grows, vanishes, and the average of stays in a ball whose radius depends on the number of agents’ feedback made available to perform (S0) in Algorithm 1 and, specifically, on the learning strategy . Next, we analyze the bound above under the lens of different learning procedures.
IvC Specifying the learning strategy
By requiring that the reconstruction error is bounded in probability, Assumption 1 is quite general and it holds true under standard assumptions for ls and gp approaches to learning . In particular, we have the following:

In parametric learning, if is modelled as an affine function of the learning parameters ’s, then setting up an ls approach to minimize the loss between the model parameters and the agents’ feedback leads to a convex quadratic program. Due to the largescale properties of ls (under standard assumptions), the error term
behaves as a normal distribution, for which Assumption
1 holds true (see [notarnicola2020distributed, Lemma A.4]), and . 
In nonparametric learning, suppose is a sample path of a gp with zero mean and a certain kernel. Due to the largescale property of such regressor and under standard assumptions, also in this case Assumption 1 holds true (see [simonetto2019personalized]) and .
Note that, in general, . Therefore, since , for the cases above we obtain thus recovering the results obtained for the perfect reconstruction case shown in §IVA.
V The timevarying case
We now investigate the GNEP in (1) in case the local cost function of each agent varies in time, thus implying that also the function is nonstationary. Our goal is still to design the parameters defining the personalized incentives to track a timevarying GNE that minimizes the (dis)satisfaction function, i.e., some , both in case of perfect (§VA) and inexact reconstruction (§VB) of the pseudogradient mapping.
To start, we make the following typical assumptions in the literature on online optimization [jadbabaie2015online, shahrampour2017distributed, dall2020optimization].
Assumption 2.
For all and , it holds that

, for ;

For all , , for ;

Moreover, , for .
Assumptions 2 i) and ii) essentially bound the variation in time of both the unknown function and the pseudogradient mappings, while Assumptions 2 iii) guarantees the boundedness of the distance between two consecutive minima such that and . Note that, with these standard assumptions in place, an asymptotic error term of the form of is inevitable [jadbabaie2015online, mokhtari2016online, NaLi2020, dall2020optimization].
Lemma 5.
Let Assumption 2 ii) hold true. For all , , with .
Va Online perfect reconstruction of the pseudogradients
In case the learning procedure allows for , for all and , we have the following ancillary results.
Lemma 6.
Let Assumption 2 ii) hold true, and , for all . Then, with the personalized incentives in , for all we have
(7) 
As in the stationary case in §IVB, the vector is not guaranteed to be a descent direction for the unknown mapping in the sense of Lemma 3. In fact, the error , introduced because of the timevarying nature of the pseudogradients, excludes that the LHS in (5) is strictly negative. The following bound characterizes the sequence of vGNE originating from Algorithm 1 in case allows for a perfect reconstruction of the timevarying mapping .
Theorem 2.
Theorem 2 says that the average of the residual over the horizon is bounded by the sum of two terms, which depend on the initial suboptimality of a computed vGNE compared to a minimum for the unknown function , and several bounds on the variations in time of , and constrained minima postulated in Assumption 2 and Lemma 5. In this case, the coordinator may reduce the error in the RHS by properly tuning the product close to one, thus leading to a large , and hence possibly boosting the convergence of Algorithm 1. In fact, if the parameter is fixed in time, we obtain This inequality ensures that will be always contained into a ball of constant radius, whose value can be adjusted through and .
VB Inexact estimate of the pseudogradients
As in §IVB, we consider the case in which, due to possibly noisy agents’ feedback, the learning procedure does not allow a perfect reconstruction of each timevarying gradient , . First, we postulate the timevarying counterpart of Assumption 1, and then we provide a preliminary result.
Assumption 3.
For all and , , and, for any , there exists and available agents’ feedback such that , for some nonincreasing function such that , for all .
Lemma 7.
Along the same line drawn for the stationary case with inexact reconstruction, we now provide the following bound on the sequence of vGNE, , generated by Algorithm 1. Note the slight abuse of notation in defining , which is different from the one in Theorem 1.
Theorem 3.
Let Assumption 2 and 3 hold true for some fixed , and , for all . Moreover, let some be fixed, and, for any global minimizer , . Then, with the personalized incentives in (4), the sequence of vGNE , generated by Algorithm 1, satisfies the following relation
(10) 
with probability , where , and is the number of agents’ feedback at the th outer iteration.
Also in this case, the average of the residual over the horizon is bounded by the sum of two terms, which depend, among the others, on the reconstruction error of the mapping and its variations in time. We note that the bound in the RHS of (4) can be adjusted through an accurate choice of the gain the stepsize . Specifically, choosing a small reduces the reconstruction error, hidden in the variable , while setting close to one induces a large value for (and for as well), thus possibly eliminating the second term under the square root of (10), and the one outside.
For simplicity, let us now suppose that the parameters and of the personalized incentives in (4) are fixed in time, namely is a constant term. From (10), we note that Due to the timevarying nature of the problem in question, also in this case the average residual can not vanish as grows, albeit the radius of the error ball can be reduced through a fine tuning of and .
VC Specifying the timevarying learning strategy
In a timevarying setting, one cannot expect to vanish in general, since the time variations in are not supposed to be asymptotically vanishing [jadbabaie2015online, dall2020optimization]. Popular learning approaches include ls with forgetting factors [Mateos09giannakis] and timevarying gp [Bogunovic2016], for which we have .
Vi Ridehailing with maas orchestration
We verify our findings by resuming the motivating example in §IA, and then running simulations on a numerical instance.
Via Problem description
The parameters adopted are retrieved from open data collected in New York City in April 2019 [nyc_ridehailing] that provide information on companies: Yellow taxi, Uber, Lyft, Juno and Via. We stress that the of the ridehailing service, measured as the total number of vehicles deployed on the road, coincides with an integer variable, i.e., , for all , thus leading to a mixedinteger setting. However, since the fleet dimension of each firm we are considering is in the order of few thousands of vehicles (i.e., Juno and Via), or tens of thousands for bigger companies (Uber, Lyft, Yellow taxi), we consider the relaxed version associated to the example in §IA by treating as a scalar continuous variable, and then rounding its value [pandey2019needs]. For this reason, we roughly estimate a roundoff error in the order of for any GNE computed at item (S2) in Algorithm 1 through an extragradient type method [solodov1996modified] (thus partially neglecting the multiagent nature of the inner loop). Moreover, we decide to split the hours of a day in intervals, enumerated in the set , according to the estimated average travel time of each costumer with no shared trips, i.e., about minutes. Then, at every , each firm aims at solving the following mutually coupled optimization problem
(11) 
where denotes the number of cars deployed by company , lower and upper bounded by ,