This article introduces an argumentation-based decision process framework speci cally designed to model context-based decisions and its application to the challenge of promoting responsible modal choices in transportation. Despite the growing demand for sustainable transportation options, many urban travelers continue to rely heavily on private cars. We show that our argumentation model can be used to understand how the traveler context in uences the transportation modal choice decisions of the travelers. To validate the e cacy of our framework, we deploy it within a simulator of multimodal transportation networks, utilizing formal argumentation to represent various behaviors. By examining the underlying reasons behind individuals' car usage and investigating potential avenues for in uencing their modal choices, we aim to contribute to the advancement of sustainable transportation solutions.
The growth of cities is accompanied by an increasing transportation demand, resulting in heightened pollution and congestion. This is primarily attributed to travelers’ preference for using private vehicles over other modes of transportation. The shift from private vehicle mode to alternative modes such as collective or non-motorized modes has become a signi cant concern for transportation authorities. To discourage private vehicle usage, authorities have implemented low-emission zones (LEZ) as a new measure. LEZ restricts access to certain parts of the city exclusively to vehicles with low emissions.
However, de ning the boundaries of these zones is a challenge as it requires striking a balance between travelers’ mobility needs and the tra c implications. Unfortunately, when the de nition is solely based on tra c ow analysis, only the tra c consequences are taken into account. This approach is unfair as it places the burden solely on excluded travelers or those who can a ord low-emission vehicles.
Neglecting the impact on travelers’ needs presents two risks. Firstly, there is a high likelihood of non-compliance, resulting in additional costs to enforce the rule. Secondly, there is a limited e ect on modal shift, as most travelers simply adjust their car routes to avoid the LEZ. Finally, to align with the users’ needs, cities may introduce exceptions that make the rule unclear12. The LEZ de nition problem emphasizes the necessity of conducting a more comprehensive analysis of travelers’ decision-making processes to grasp the rule’s impact on travelers and, consequently, assess its e ectiveness.
Agent-based simulations focus on individual decision-making processes, making them
valuable tools for analyzing the modal shift problem [
Modal shifting is determined by a whole range of factors that are interrelated to a larger or
smaller extent. For instance, [
To address this complexity, this article aims to propose a framework that captures the various contexts within which travelers make their travel decisions.
In this article, we argue that formal argumentation, such as the Dung framework [
The contributions of this article are: • An argumentation framework to represent context-based decisions, • An application of argumentation to the problematic of the modal shift.
This paper is organized as followed: Firstly, Section 2 proposes a state of the art on argumentationbased decisions frameworks. Secondly, Section 3 introduces the case study dedicated to represent urban travelers behaviors within a simulator of multimodal transportation networks. Thirdly, Section 4 presents the formal framework and recalls basics notions. Finally, Section 5 presents the rst results of the proof-of-concepts and Section 6 proposes conclusion and perspective of future work.
1https://ec.europa.eu/transport/themes/urban/studies_cs 2https://ec.europa.eu/transport/sites/default/ les/uvar_ nal_report_august_28.pdf
Argumentation has been identi ed as an e ective tool for decision-making and decision-support
systems, particularly in situations where the recommended decisions need to be explained [
In this article, we are interested in the model proposed in [
Value-based Argumentation Frameworks (VAF) [
The proposal presented in this paper is evaluated using a multiagent simulator available in the Plateforme Territoire3. This simulator enables agent travelers to access multimodal shortest itineraries between their origin and destination and simulates their movement along the chosen itinerary at a speed corresponding to the selected transportation mode. Itineraries can be evaluated using pre-trip indicators that in uence the itinerary choice of the traveler agents. These indicators can be based on factors such as distance or tra c-related aspects like noise, which depends on the number of vehicles along di erent parts of the itinerary. Additionally, the simulator calculates global tra c indicators for each transportation mode to assess the system, such as the number of late travelers.
Application. Each agent has to decide about one alternative which corresponds to choose a particular transportation network. In this simulator, we consider the following set of alternatives Alts and N be a set of agents :
D1 p.t. :="go by public transport" D2 bike :="go by bike" D3 walk :="go by foot" D4 car :="go by car"
Each agent decides based on indicators. For each agent i N , we consider the following indicators Inds. We rst de ne the indicators based on alternatives Alts: • t : Alts • d : Alts • pol : Alts
alternative • noi : Alts • cos : Alts
D+ for a given alternative, it returns the duration for this alternative D+ for a given alternative, it returns the distance for this alternative
D+ for a given alternative, it returns the pollution rate associated with this D+ for a given alternative, it returns the noise generated by this alternative
D+ for a given alternative, it returns the cost of this alternative Indicators based on the agent state: • em : N { , }
is a function that represents if one agent has a medical emergency • isOld, isF emale, isReadyT oM odalShif t : N { if one agent is old, is female, or is ready to modal shift4 , } are functions that represent • hasCar, hasECar, hasBike : N { , } are functions that represent if one agent has a car, or has an electric car, or has a bike and are s.t. for each agent i N , if hasCar(i) = then hasECar(i) =
• isH ealthCrisis, isRushH our, isT heN ight { if this is the rush hour, or if it is the night , } translate if there is a health crisis, We formally de ne the state space (based on previously de ned indicators) such as : SpInds = ((D+)Alts N )4 ({ , } )Alts N ({ , }
N )7 { , } 3
s S pInds, ind I nds, s[ind] = ind
This last notation translates for all s S pInds, s[t] = t i.e. we return the part of the value of the component of vector that assigns the function which evaluates the duration of each alternative.
4For a sake of simplicity, we reduce to a small set of characteristics of the agent.
In this section we give the model of Amgoud and Prade [
In order to map arguments to decisions, [
De nition 1. An Argumentation Framework for Decision Making (AFDM) is a tuplet AF DM = (A, R, D, Ff , Fc) such that: • A is a set of arguments • R is a binary relation called attack relation • D is a set of decisions (or actions) • Ff : D 2A is a function that assigns from D the set of pro arguments • Fc : D 2A is a function that assigns from D the set of con arguments
We note ADF (A, D) = 2A 2A A 2D (2A)D (2A)D the set of Argumentation-based Decision Framework based on a set of arguments A and a set of decisions D and consider: a A , Ff 1(a) := {d D : a F f (d)}, Fc 1(a) := {d D : a F c(d)}
A semantics of argumentation frameworks is given by the notion of extensions [
De nition 2 (Extensions). Let (A, R, D, Ff , Fc) be an AFDM, S A be a set of arguments. • S is an admissible extension i S is con ict-free and all arguments A S are acceptable w.r.t. S. • S is a complete extension i S is admissible and contains all acceptable arguments wrt S. • S is a grounded extension i S is a minimal complete extension wrt i.e. S A s. t.
S S, S is a complete extension. • S is a preferred extension i S is a maximal admissible extension wrt . • S is a stable extension i S is con ict-free, and A A \ S, SRA. • S is an ideal extension i S is a maximal admissible extension wrt that is included in all preferred extensions.
Let us notice that in the general case there is no consensus about which extension semantics
to use. However as suggested in [
The Argumentation-based Decision Process (ADP) framework incorporates argumentation theory to model the decision-making process, enabling the selection of decisions based on argument extensions derived from the argumentation graph.
Thus, the system is fully described by a set of agents N , a set of states S, a set of arguments A and a set of decisions D or actions that agents can do. In the sequel, we de ne an argumentationbased decision process for one agent i N . A function de nes for each state s S , an instance of Amgoud and Prade’s model i.e. a subset of possible decisions D D that in s the agent can do, a subset of arguments A A which are veri ed in s, a set of (un)supported decisions Ff (Fc), and a set of attacks R that are generated by the semantics of each argument in A . Then, a function de nes the extension semantics which is used by the agent to compute her stationary politics . Since there is no concensus about how to compute the "winning" set of arguments based on a particular extension semantics (and so the decision), we let abstract and consider rather an heuristic function h which de nes the computation method to choose a decision based on an extension semantics. Thus, we assume that the politic of the agent (which is stationary) is fully de ned by this heuristic function i.e. = h. The stationary policy function ensures that a decision is chosen from the available options for each state in a consistent manner w.r.t. the set of admissible arguments.
De nition 3. An Argumentation-based Decision Process (ADP) is a tuplet ADP = (S, A, D, , , ) such that : • S is a no-empty set of states • A is a no-empty set of arguments • D is a no-empty set of decisions (or actions) • : S ADF (A, D) is a function s.t.
s S , (s) = (A , R , D , Ff , Fc) where : – R
A A [R](s) = R – A A is a subset of arguments associated with state s and we note [A](s) = A represents the set of attacks between arguments in A and we note •
: 2A A extensions – D D is a subset of decisions associated in a state s and we note [D](s) = D – Ff : D assigns a set of pro arguments for each decision in D in a state s
2A and we note [Ff ](s) = Ff and [Ff 1](s) = Ff 1 is a function that – Fc : D 2A and we note [Fc](s) = Fc and [Fc 1](s) = Fc 1 is a function that assigns a set of con arguments for each decision in D 22A is a function that, from an argumentation graph, returns the set of
: S D is a stationary politic s.t. s S , (s) = hs( ( [R](s))) where hs : 22A D is a function s.t. each chosen decision belongs to the set of extensions given by the AFDM: E 2A, hs(E ) { d D : d [D](s)}
We now present how this model is applied in our proof-of-concept. The implemented model has 22 arguments, and for a sake of readability, we do not present all of these arguments in this article but only 4 of them.
Application. Let consider the following ADP = (S, A, D, , , ) where S = SpInds and D = Alts. We consider a set of arguments A and we assume in our study case that is such that for all s S and for each agent i N : • A. hC := "agent i has no car" ; C. hB := "agent i has no bike",
Activation : hC and hB
[A](s) i s[hasCar](i) = Pros : [Ff 1](s)(hC) = [Ff 1](s)(hB) = Cons : [Fc 1](s)(hC) = [Fc 1](s)(hB) = [A](s) i s[hasBike](i) = Attacks : (a, d) { (hB, bike), (hC, car)}, {(a, x) : x [Ff ](s)(d)} [R](s) Meaning : If she has no car (resp. no bike), then hC (resp. hB) is veri ed. It is not in favor, or against any alternative and attacks all veri ed arguments that supports one of these alternatives. • iAE := "it is a medical emergency"
Activation : iAE Pros : [Ff 1](s)(iAE) = {d [A](s) i s[em](i) = [D](s) : ¬ x
[D](s), s[t](i)(x) > s[t](i)(d)} Cons : [Fc 1](s)(iAE) = [D](s) \ [Ff 1](s)(iAE) Attacks : {(iAE, x) : x [A](s), [Ff 1](s)(x) [Fc 1](s)(iAE) = } [R](s) Meaning : If there is a medical emergency, then iAE is veri ed. It is in favor of all alternatives that are the quickest, against the others and attacks all veri ed arguments that are in favor of at least one alternative that is not the quickest. • cRA := "the car alternative crosses the regulated area"
Activation : cRA [A](s) i s[reg](i)(car) = , s[hasCar](i) = and car [D](s) Pros : [Ff 1](s)(cRA) = Cons : [Fc 1](s)(cRA) = {car} Attacks : {(cRA, x) : x [Ff ](s)(car)}
[R](s) Meaning : cRA is veri ed when the alternative car crosses a regulated area while it should be forbidden. It attacks all arguments in favor of car.
[A](s) i • iEx := "it is too much expensive for agent i when s[cos](i)(d) > ic(s)(d)" with d D and ic(s) : D D+ a function to set a threshold for what the agent considers as too much expensive Activation : iEx d
[D](s) s.t. cos(i)(s)(d) > ic(s)(d) Pros : [Ff 1](s)(iEx) = {d Cons : [Fc 1](s)(iEx) = {d Attacks : {(iEx, x) : d [D](s) : ic(s)(d) > s[cos](i)(d))} [D](s) : s[cos](i)(d) > ic(s)(d)} [Fc 1](s)(iEx), x [Ff ](s)(d)}
[R](s) Meaning : iEx is veri ed when at least one alternative is above the threshold of acceptability of agent i in regard to her budget. It attacks all arguments that supports one alternative which is not in the threshold of acceptability.
It is worth noting that attacks hold more weight in the decision-making process compared to simply considering a list of pros and cons. This is because a single attack can invalidate an argument and subsequently remove it from the extensions, thus impacting the agent’s deliberation process for making choices.
Furthermore, this model could be easily extended to get a more realistic models by considering other arguments as e.g. "I’m relocating", "I’m the police", "I prefer biking", "There is no bicycle network", "There is no bus at this hour", etc.
In this section, we provide an application of the framework to simulate the modal shifting. We present the results obtained from our experiments, starting with an explanation of the various scenarios tested. Then, we provide an example of an argumentation graph generated for one agent. Finally, we demonstrate how decision-making based on argumentation can be utilized to evaluate the evolution of the overall transportation network.
We aim to enhance the simulation of regulating a multimodal transportation networks, consisting of a car, bus, walk, and bicycle network, each with distinct characteristics such as average speed, environmental impact, nancial cost, and noise level. Speci cally, we focus on regulating the car network, which involves determining certain areas where cars are either allowed or prohibited from accessing. We present the parameters considered in two network con gurations: namely, a con guration without regulated areas and another with regulated areas (LEZ). In both, we assume there are 5000 traveler agents, no health crisis, and it is rush hour. The objective of a traveler agent is to choose the most appropriate transportation mode. We also establish some thresholds based on these assumptions e.g. 90% have a bike and 100% a car, 30% of them have an electric car, 40% are ready to modal shift, 20% are senior, 10% are emergencies. The threshold for the acceptability w.r.t. time i.e. it(s)(d) is set arbitrary as the following : if the travel time is greater than 1.2 tmin then it is unacceptable. The noise threshold in(s)(d) should not exceed 1.0. The quantity of pollution, ip should not exceed 9.0, the cost becomes unacceptable when the cost is greater than ic(s)(d) = 0.2 and the distance is unacceptable when it is greater than il = 200.0. We consider 6 scenarios: • Without LEZ: the objective is to evaluate the impact of taking into account the context in the distribution of the travelers per networks considering a classical multimodal network. – s1: the traveler agents decide by considering the quickest alternative, i.e. a monocriteria decision process. – s2: the traveler agents decide with the proposed ADP. • With LEZ: the objective is to evaluate if our model is e cient to understand the traveler decision process for the transportation modal shift.
– s3: the traveler agents choose the quickest alternative. The comparison with s1 will give information about the quality of the decision criteria to understand the consequences of the LEZ de nition. – s4: the agents decide with an ADP. The comparison with s2 will give information about the quality of the decision criteria to understand the consequences of the LEZ de nition. – s5: the traveler agents decide with the ADP de ned in s4 but the distance acceptance has been reduced. The comparison with s4 should show an increase of the modes with the shortest distance. – s6: the traveler agents decide with the ADP de ned in s5 but the pollution acceptance has been reduced. The comparison with s5 should show an increase of the least polluting modes.
After running the scenario s4 (ADP+LEZ), we illustrate an argumentation graph of one agent. Its characteristics are given by : s[em](i) = , s[isOld](i) = , s[isRT MS](i) = , s[hasCar](i) = , s[hasECar](i) =
, s[hasBike](i) = .
Application. If Grd( [R](s)) is the computed grounded extension from the argumentation graph [R](s) where s represents the current state, then, we de ne the scoring function scr1 : [D](s) R such that for all decisions d [D](s) : pros(d) = |Grd( [R](s)) cons(d) = |Grd( [R](s)) [Ff ](s)(d)| [Fc](s)(d)| scr (d) =
pros(d) pros(d)+cons(d) 0 otherwise
if pros(d) + cons(d) = 0
We compute the scores by considering the argumentation model given in Application 4.2. The result of the computed argumentation graph is depicted in Table 2. Then, by computing the grounded extension, we get: Grd( [R](s)) = {{5, 17}}. To deal with equalities, we assume the following order for agent i: Bus > Bike > W alk > Car. For a sake of simplicity we decided to set this order arbitrary. In this setting, the agent chooses the car alternative by avoiding the regulated area since scr(Car) = 1 while for other alternatives X, we have scr(X) = 0.
We analyze the impact of a LEZ on agents considering scenarios s1, s2 (ADP), s3 (LEZ), s4 (ADP+LEZ), s5 (ADP+LEZ), s6 (ADP+LEZ). The Table 1 presents the distribution of the travelers between the transportation modes according to several scenarios. Here we compare the scenarios to illustrate three advantages of our approach. In each of them, travelers have to choice the transportation modes corresponding to their preferences with or without considering LEZ.
Scenario Is our approach adapted to reproduce the diversity of travelers’ modal choices? The scenario s1 considers that travelers choose the quickest trip without other parameters while the travelers in s2 decide according to the argumentation model presented in section 5.1. Here, the result show that with the only decision criteria based on time it is not possible in this example to have a real multimodal system, the car alternative is always the fastest transportation mode. The argumentation model is closer to the reality.
Is our approach e icient to understand the consequences of the network regulation on multimodal travelers’ modal choices? The scenarios s3 and s4 are respectively similar to s1 and s2 expected that a LEZ is deployed. We can observe for both the same evolution with the transfer of around 5% of travelers from the car mode to the bus mode. The advantage of our proposal is that this transfer being based on a more realistic initial distribution we observe a multimodal tra c that is more balanced between modes.
Is our approach adapted to understand the consequences of traveler behaviors on the transportation system? The scenario s5 is based on s4 (LEZ deployed and ADP) with the modi cation of the distance constraint argument ( il) which is reduced to 0. It means that travelers prefer the shortest trips. We observe that there is a shift of travelers towards the car mode to reach a value close to that of s2. This illustrates that the decision is the result of a compromise, as the majority of travelers do not shift.
The scenario s6 is based on s5 with the reduction of the tolerance of the pollution ( ip). This argument counter balances partially the one of the distance. The result is a percentage of travelers choosing the car mode that is similar to s4 and an increase of the traveler choosing the bus mode. This last mode is a good compromise between the distance and pollution arguments.
In this article, we have presented an argumentation-based decision process framework for modeling the decision-making process of urban travelers. To demonstrate the feasibility of our framework, we have provided a proof-of-concept by instantiating the model with arguments that could be considered in a modal shifting.
It is important to highlight that our current model lacks of realistic data on real behavior of urban travelers, and we acknowledge that it represents a preliminary e ort in this direction. In future research, we plan to enhance our framework by incorporating e.g. web-based data and considering more statistical data. We believe that it will provide more accurate insights into the decision-making process of agents and will provide more realistic results.
By combining our framework with comprehensive and up-to-date web data, we anticipate that our model will o er a deeper understanding of urban travelers’ decision processes, ultimately leading to more e ective strategies for promoting responsible modal choices in transportation.
We thank the 3IA Côte d’Azur ANR-19-P3IA-0002, the HyperAgents project ANR-19-CE23-0030 and the Acceler-AI project Projet-ANR-22-CE23-0028 for their support. D {
} } } Fc D{ ,D {} ,3D {} 4D ,4 ,4 D { D {} ,3D
D D 4 } 4 { { { } } 1 1 D D { { } 2 D { } 2 } 2 } } } } } } } 2 1 D 2 2 2 2 D 1
} } } , , 1 D 4 3 D 2 3 D D D D 3 F 4 D D 4 D D 3 3 3 3 D {} ,4D f , , , , , ,
{ { { , , D D 1 D D D D 1 D { { { { { { {
D {
D { } 1 1 , 2 } 1} 1 1 , ,1 ,1 0} 0} 01 2 2 1 1 , s ,1 ,1 ,7 ,7 ,6 ck 0 ,6 ,6 ,6 13 a {} ,61 } 3 } } 1 1 } } } , } t { 1 { { , , { { { 7 { t , , 0 0 , A ,7 ,9 ,2 ,2 ,9 9 6 1 1 6 , 1 , , 1 6 , 5 5 , 1 5 { { 0 { 1{ 2 , 5 1 , 1 { s i s i r D I g 1 5 6 7 9 10 11 12 13 15 16 17 20 r
A