Prise de décisions sous incertitude sur la base de données: conscience, apprentissage et cohérence interpersonnelle – DATA-AWARE

In this project, we propose a model in which information about possible contingencies, and about their probabilities is obtained from available data as in the case-based decision theory. Data are intrinsically incomplete and provide only partial information about the underlying state-space and the relevant probabilities. The evaluation of actions thus has to combine the objective information with subjective evaluation of uncertainty: ambiguity (uncertain probabilities over outcomes), and unawareness (uncertainty about the relevant states).
The model allows for different types of learning: Bayesian updating; learning about new states; refining the state space; and, learning about counterfactuals. Such learning leads to a revision of the subjective model of uncertainty; may lead to updating of the perception of uncertainty; and allows for adjustments of the attitudes towards uncertainty. An important question concerns the acquisition and evaluation of new information. We propose a concept for determining the value of information and relate it to the decision maker’s subjective attitude towards ambiguity and unawareness.

We developed a framework of decision making under uncertainty where the main source of information about possible outcomes of actions and the circumstances determining them is based on sets of data about recorded cases. Two main issues evolved from the research. First, the information from the data set can be represented as a mass distribution over sets of possible outcome distributions. This allowed us to derive an axiomatic representation of preferences based on data and study a special case based on NEO-additive capacities. Second, the possibility of active search for data specifying either the conditions for application of particular actions, or the existence of new relevant characteristics. Notably, we introduce measurement actions which capture the active search for new data.
The introduction of measurement actions links the data-based approach to the problem of information acquisition and the value of information. The latter was investigated in co-operation with Illia Pasichnichenko (University of Sussex) for situations in which the signal generating process and the posterior distributions are only partially known. The value of information is thus related to the decision maker's attitude towards ambiguity. This work introduces a new concept of informativeness, proves a theorem in the spirit of Blackwell, and characterizes the value of information in terms of the preference relation over information structures. Depending on ambiguity attitude, the value of information may be negative.
In a collaboration with John Quiggin, (Queensland University) and Evan Piermont (Royal Holloway), we analyse how agents who have distinct perception of the underlying uncertainty and express it by means of distinct languages (syntax) can communicate via translation. We identify conditions under which the individual state spaces and languages can be embedded into a common state space / joint language with respect to which each individual representation is a “small world”. We also highlight situations in which no such common state space exists.
A collaboration with Simon Grant (ANU) and David Kelsey (University of Nottingham) studies the problem of dynamically consistent equilibria in sequential games, when there is ambiguity about opponents' behaviour. It developed a concept of consistent-planning equilibrium and applied it to sequential games. Notably, moderate degree of optimism coupled with sufficient ambiguity produces behavior typically observed in experimental studies.

We plan to continue the study of value of information by integrating bounbded awareness, as well as the trade-off between exploration and exploitation. The results on communication between agents with distinct awareness and on strategic behaviour under uncertainty will be used to explore the emergence of joint language and a common state-space in dynamic cheap-talk games. Finally, we will also study the relationship between human and machine learning with possible applications to financial markets.

Discussion Papers:

Eichberger, J. , Grant, S., Kelsey, D. (2025). «Ambiguity in Multi-Stage Games: Uncertainty about the Opponent’s Strategic Behavior«

Eichberger, J. , Guerdjikova, A.(2024). «Data, Cases and States«

Guerdjikova, A., Piermont, E., Quiggin, J. (2025). «Do You Know What I Mean? A Syntactic Representation for Differential Bounded Awareness«

Seminars and Conference Presentations:
1. Jürgen Eichberger: Invited presentation at the Workshop on Unawareness, University of Aarhus, Aarhus, Demark: “Data, Cases and States”, May 30 – June 1, 2024
2. Ani Gierdjikova: Invited presentation at the Workshop on Unawareness, University of Aarhus, Aarhus, Demark: “Do You Know What I mean: A Syntactic Representation for Differential Bounded Awareness”, May 30 – June 1, 2024
3. Jürgen Eichberger: FUR 2024, Brisbane “Data, Cases and States”, July 4 – July 7, 2024
4. Ani Guerdjikova: FUR 2024, Brisbane “Do You Know What I mean: A Syntactic Representation for Differential Bounded Awareness”, July 4 – July 7, 2024
5. Jürgen Eichberger: Invited seminar, ANU, Canberra “Data, Cases and States”, August 8, 2024
6. Ani Guerdjikova: Invited seminar at the University of Manchester: “Do You Know What I mean: A Syntactic Representation for Differential Bounded Awareness”, October 16, 2024
7. Ani Guerdjikova: Presentation at the meeting of the editorial board of the Journal of Mathematical Economics, : “Do You Know What I mean: A Syntactic Representation for Differential Bounded Awareness” December 16, 2024
8. Ani Guerdjikova: Presentation at SAET, Ischia: “Data, Cases and States”, June 29 – July 6, 2025
9. Ani Guerdjikova: Invited seminar at Heidelberg University, “How Do You Know What I Mean?”, October 21, 2025
10. Ani Guerdjikova: Invited seminar at Bielefeld University, “How Do You Know What I Mean?”, October 28, 2025
11. Jürgen Eichgberger: Presentation at the workshop Unforeseen Contingencies: Representations and Applications, June 5-6, 2025 in Grenoble: “Data, Cases and States”

Le modèle standard de prise de décision sous incertitude est basé sur l’espace des états du monde: exogène, objectif, observable et commun à tous les agents. Les concepts d’arbitrage et de dominance, fondamentaux dans le discours économique, sont ancrés dans cette hypothèse. Comment les agents économiques construisent-ils cet espace? Et comment arrivent-ils à une connaissance commune de celui-ci?
Nous proposons un modèle, dans lequel l’information concernant les contingences possibles, ainsi que leurs probabilités, est disponible sous forme de données, comme dans la théorie des décisions à cas-par-cas. Les données sont intrinsèquement incomplètes. L’information concernant l’espace des états du monde est donc partielle. L’évaluation des actions doit combiner les informations objectives avec une évaluation subjective de l’incertitude sous-jacente: ambiguïté (probabilités inconnues), et conscience limitée (ensemble des états du monde inconnu).
Ce modèle génère plusieurs types d’apprentissage: apprentissage Bayésien; apprentissage des nouveaux états du monde; affinement des états du monde connus; apprentissage des contrefactuels. Cet apprentissage induit une révision du modèle subjectif d’incertitude, mais aussi de la perception d’incertitude et de l’attitude envers celle-ci.
Les caractéristiques subjectives des décideurs influencent leur modèle d’incertitude. Alors, des individus ayant accès à des données identiques peuvent différer par rapport à leurs prévisions subjectives et à leurs comportements. Une telle hétérogénéité a des implications observables pour les comportements économiques, les allocations du marché, les interactions stratégiques et les politiques publiques, au-delà du contenu statistique des données.
Le modèle proposé est ensuite utilisé pour étudier des applications économiques pertinentes. Nous nous intéressons à la question, si une société composée des individus ayant accès à la même base de donnée arrivera à un modèle commun de l’incertitude et atteindra connaissance commune de l’ensemble des états du monde et des probabilités sous-jacentes. Ensuite, nous modélisons des interactions stratégiques, dans lesquelles les croyances d’équilibres sont influencées par des données sur les choix passés des joueurs. Nous en étudions les conséquences pour les comportements stratégiques à long-terme et l’émergence des institutions. Notre modèle constitue un homologue subjectif aux modèles d’apprentissage statistique et de l’IA, basés purement sur des critères objectifs. Nous explorons ses implications pour le problème classique de classification. Finalement, nous appliquons le modèle à l’analyse des comportements, des allocations d’équilibre et des prix dans les marchés financiers. Le but est d’expliquer les faits empiriques et de comprendre les conséquences du long-terme des décisions informées par des données par rapport à l’hétérogénéité des caractéristiques individuelles et des croyances.