ANR-MIAI in contracts (2024-02-10)

Reproducibility is a desirable property of scientific research. On the one hand, it increases confidence in results. On the other hand, reproducible results can be extended on a solid basis. In rapidly developing fields such as machine learning, the latter is particularly important to ensure the reliability of research. In this paper, we present a systematic approach to reproducing (using the available implementation), replicating (using an alternative implementation) and reevaluating (using different datasets) state-of-the-art experiments. This approach enables the early detection and correction of deficiencies and thus the development of more robust and transparent machine learning methods. We detail the independent reproduction, replication, and reevaluation of the initially published experiments with a method that we want to extend. For each step, we identify issues and draw lessons learned. We further discuss solutions that have proven effective in overcoming the encountered problems. This work can serve as a guide for further reproducibility studies and generally improve reproducibility in machine learning.

Artificial agents, as humans, use their knowledge to behave in an environment and within a society. Humans evolve their knowledge by adapting it in response to interactions with their environment and society. The question that is raised in this thesis is:``can knowledge evolve in a society of artificial agents, as it does in a human society?'' In particular, if agents adapt to improve their social interactions, how can this affect the quality of the population's knowledge about the environment? And how does it affect knowledge diversity? To address the questions, ontology evolution is simulated based on principles from experimental cultural evolution through an experimental framework in which: agents initially learn ontologies, from object samples, which they later adapt by interacting with each other about objects in the environment. Using this experimental framework, we show that: (1) agents reach a state of agreement in their interactions, (2) they improve the quality of their knowledge about the environment, and (3) they preserve the diversity of their knowledge. In order to characterise knowledge evolution through multiple generations, experiments are conducted with agents endowed with reproduction capabilities. Results show that (1) the variation provided by inter-generation transmission allows agents to further improve the quality of their ontologies; (2) agents select the knowledge to be preserved through intra-generation transmission which compensates for the lack of teacher selection in inter-generation transmission; and finally, (3) diversity remains stable from one generation to another. This work not only provides a basis for implementing agents capable of culturally evolving their knowledge, but also suggests that simulating such behavior can serve as a valuable tool for testing hypotheses about human cultural knowledge evolution.

Des agents peuvent faire évoluer leurs ontologies en accomplissant conjointement une tâche. Nous considérons un ensemble de tâches dont chaque agent ne considère qu'une partie. Nous supposons que moins un agent considère de tâches, plus la précision de sa meilleure tâche sera élevée. Pour le vérifier, nous simulons différentes populations considérant un nombre de tâches croissant. De manière contre-intuitive, l'hypothèse n'est pas vérifiée. D'une part, lorsque les agents ont une mémoire illimitée, plus un agent considère de tâches, plus il est précis. D'autre part, lorsque les agents ont une mémoire limitée, les objectifs de maximiser la précision de leur meilleures tâches et de s'accorder entre eux sont mutuellement exclusifs. Lorsque les sociétés favorisent la spécialisation, les agents n'améliorent pas leur précision. Cependant, ces agents décideront plus souvent en fonction de leurs meilleures tâches, améliorant ainsi la performance de leur société.

Agents have been previously shown to evolve their ontologies while interacting over a single task. However, little is known about how interacting over several tasks affects the accuracy of agent ontologies. Is knowledge learned by tackling one task beneficial for another task? We hypothesize that multi-tasking agents tackling tasks that rely on the same properties, are more accurate than multi-tasking agents tackling tasks that rely on different properties. We test this hypothesis by varying two parameters. The first parameter is the number of tasks assigned to the agents. The second parameter is the number of common properties among these tasks. Results show that when deciding for different tasks relies on the same properties, multi-tasking agents reach higher accuracy. This suggests that when agents tackle several tasks, it is possible to transfer knowledge from one task to another.

Cultural values are cognitive representations of general objectives, such as independence or mastery, that people use to distinguish whether something is "good" or "bad". More specifically, people may use their values to evaluate alternatives and pick the most compatible one. Cultural values have been previously used in artificial societies of agents with the purpose of replicating and predicting human behavior. However, to the best of our knowledge, they have never been used in the context of cultural knowledge evolution. We consider cooperating agents which adapt their individually learned ontologies by interacting with each other to agree. When two agents disagree during an interaction, one of them needs to adapt its ontology. We use the cultural values of independence, novelty, authority and mastery to influence the choice of which agent adapts in a population of agents sharing the same values. We investigate the effects the choice of cultural values has on the knowledge obtained. Our results show that agents do not improve the accuracy of their knowledge without using the mastery value. Under certain conditions, independence causes the agents to converge to successful interactions faster, and novelty increases knowledge diversity, but both effects come with a large reduction in accuracy. We however did not find any significant effects of authority.

In cultural knowledge evolution simulated by multi-agent simulations, agents can improve the accuracy of their knowledge by interacting with other agents and adapting their knowledge with the aim of agreeing. But their knowledge might be confined to specific areas because they do not have the capacity to explore the world around them. Since intrinsic motivation to explore in artificial agents has already proven to increase exploration, it was researched whether and how agents in simulations of cultural knowledge evolution can be motivated to explore. Moreover, it was tested how far this improves and changes their knowledge. Three different kinds of motivation were investigated: curiosity, creativity and non-exploration. Moreover, intrinsic motivation was modelled with and without reinforcement learning. Agents either explored on their own or picked specific interaction partner(s). It has been shown that it is possible to model agents with intrinsic motivation to explore in cultural knowledge evolution, and that this has a significant effect on the agents’ knowledge. Contrary to the expectations and other studies, this did not lead to an increase in knowledge completeness. Out of all intrinsic motivations, curiosity had the highest accuracy and completeness. Models with reinforcement learning performed similar to direct models. As expected, intrinsic motivation led to faster convergence of the agents’ knowledge, especially with social agents. Heterogeneously motivated agents only had a higher accuracy and completeness than homogeneously motivated agents in specific cases. This thesis can be regarded as a foundation for further investigation into the role of intrinsic motivation in cultural knowledge evolution. Different forms of intrinsic motivation or different reinforcement learning techniques could be tested. Additionally, intrinsic motivation at different stages of the experiment or in different ratios, for example curious agents and agents with no motivation, could be investigated in more detail. Lastly, agents could teach other agents things they explored a lot.

Agents use their own vocabularies to reason and talk about the world. Public signature awareness is satisfied if agents are aware of the vocabularies, or signatures, used by all agents they may, eventually, interact with. Multi-agent modal logics and in particular Dynamic Epistemic Logic rely on public signature awareness for modeling information flow in multi-agent systems. However, this assumption is not desirable for dynamic and open multi-agent systems because (1) it prevents agents to use unique signatures other agents are unaware of, (2) it prevents agents to openly extend their signatures when encountering new information, and (3) it requires that all future knowledge and beliefs of agents are bounded by the current state. We propose a new semantics for awareness that enables us to drop public signature awareness. This semantics is based on partial valuation functions and weakly reflexive relations. Dynamics for raising public and private awareness are then defined in such a way as to differentiate between becoming aware of a proposition and learning its truth value. With this, we show that knowledge and beliefs are not affected through the raising operations.

Agents have been used for simulating cultural evolution and cultural evolution can be used as a model for artificial agents. Previous results have shown that horizontal, or intra-generation, knowledge transmission allows agents to improve the quality of their knowledge to a certain level. Moreover, variation generated through vertical, or inter-generation, transmission allows agents to exceed that level. Such results were obtained under specific conditions such as the drastic selection of agents allowed to transmit their knowledge, seeding the process with correct knowledge or introducing artificial noise during transmission. Here, we question the necessity of such measures and study their impact on the quality of transmitted knowledge. For that purpose, we combine the settings of two previous experiments and relax these conditions (no strong selection of teachers, no fully correct seed, no introduction of artificial noise). The rationale is that if interactions lead agents to improve their overall knowledge quality, this should be sufficient to ensure correct knowledge transmission, and that transmission mechanisms are sufficiently imperfect to produce variation. In this setting, we confirm that vertical transmission improves on horizontal transmission even without drastic selection and oriented learning. We also show that horizontal transmission is able to compensate for the lack of parent selection if it is maintained for long enough. This means that it is not necessary to take the most successful agents as teachers, neither in vertical nor horizontal transmission, to cumulatively improve knowledge.

Les agents peuvent être utilisés pour simuler l'évolution culturelle et l'évolution culturelle peut être utilisée comme modèle pour les agents artificiels. Des expériences ont montré que la transmission intragénérationnelle des connaissances permet aux agents d'en améliorer la qualité. De plus, sa transmission intergénérationnelle permet de dépasser ce niveau. Ces résultats ont été obtenus dans des conditions particulières: sélection drastique des agents transmetant leurs connaissances, initialisation avec des connaissances correctes ou introduction de bruit lors de la transmission. Afin d'étudier l'impact de ces mesures sur la qualité de la connaissance transmise, nous combinons les paramètres de deux expériences précédentes et relâchons ces conditions. Ce dispositif confirme que la transmission verticale permet d'améliorer la qualité de la connaissance obtenue par transmission horizontale même sans sélection drastique et apprentissage orienté. Il montre également qu'une transmission intragénérationnelle suffisante peut compenser l'absence de sélection parentale.

Assessing knowledge diversity may be useful for many purposes. In particular, it is necessary to measure diversity in order to understand how it arises or is preserved; it is also necessary to control it in order to measure its effects. Here we consider measuring knowledge diversity using two components: (a) a diversity measure taking advantage of (b) a knowledge difference measure. We present the general principles and various candidates for such components. We discuss how these measures may be used to generate populations of agents with controlled levels of knowledge diversity.

The content of experiments must be semantically described. This topic has already been largely covered. However, some neglected benefits of such an approach provide more arguments in favour of scientific knowledge graphs. Beyond being searchable through flat metadata, a knowledge graph of experiment descriptions may be able to provide answers to scientific and methodological questions. This includes identifying non experimented conditions or retrieving specific techniques used in experiments. In turn, this is useful for researchers as this information can be used for repurposing experiments, checking claimed results or performing meta-analyses.

When agents independently learn knowledge, such as ontologies, about their environment, it may be diverse, incorrect or incomplete. This knowledge heterogeneity could lead agents to disagree, thus hindering their cooperation. Existing approaches usually deal with this interaction problem by relating ontologies, without modifying them, or, on the contrary, by focusing on building common knowledge. Here, we consider agents adapting ontologies learned from the environment in order to agree with each other when cooperating. In this scenario, fundamental questions arise: Do they achieve successful interaction? Can this process improve knowledge correctness? Do all agents end up with the same ontology? To answer these questions, we design a two-stage experiment. First, agents learn to take decisions about the environment by classifying objects and the learned classifiers are turned into ontologies. In the second stage, agents interact with each other to agree on the decisions to take and modify their ontologies accordingly. We show that agents indeed reduce interaction failure, most of the time they improve the accuracy of their knowledge about the environment, and they do not necessarily opt for the same ontology.

Ontology alignments enable agents to communicate while preserving heterogeneity in their knowledge. Alignments may not be provided as input and should be able to evolve when communication fails or when new information contradicting the alignment is acquired. The Alignment Repair Game (ARG) has been proposed for agents to simultaneously communicate and repair their alignments through adaptation operators when communication failures occur. ARG has been evaluated experimentally and the experiments showed that agents converge towards successful communication and improve their alignments. However, whether the adaptation operators are formally correct, complete or redundant could not be established by experiments. We introduce a logical model, Dynamic Epistemic Ontology Logic (DEOL), that enables us to answer these questions. This framework allows us (1) to express the ontologies and alignments used via a faithful translation from ARG to DEOL, (2) to model the ARG adaptation operators as dynamic modalities and (3) to formally define and establish the correctness, partial redundancy and incompleteness of the adaptation operators in ARG.

To reason and talk about the world, agents may use their own distinct vocabularies, structured into knowledge representations, also called ontologies. In order to communicate, they use alignments: translations between terms of their ontologies. aHowever, ontologies may change, requiring their alignments to evolve accordingly. Experimental cultural evolution offers a framework to study the mechanisms of their knowledge evolution. It has been applied to the evolution of alignments in the Alignment Repair Game (ARG). Experiments have shown that, through ARG, agents improve their alignments and reach successful communication. Yet, these experiments are not sufficient to understand the formal properties of cultural knowledge evolution. This thesis bridges experimental cultural knowledge evolution with a theoretical model of cultural knowledge evolution in logic. This is achieved by introducing Dynamic Epistemic Ontology Logic and defining a faithful translation of ARG in DEOL that (a) encodes the ontologies, (b) maps agents' ontologies and alignments to knowledge and beliefs, and (c) captures the adaptation operators through announcements and conservative upgrades. This model shows that all but one adaptation operator are correct, they are incomplete and some are partially redundant. Three differences between the ARG agents and their logical model explain these results, leading to an independent model of awareness based on partial valuations and weakly reflexive relations. An alternative model of ARG is then defined under which the formal properties are re-examined, showing that this model is closer to the original game. This is a first step towards defining a theoretical model of cultural knowledge evolution.

Ontology alignments enable agents to communicate while preserving heterogeneity in their information. Alignments may not be provided as input and should be able to evolve when communication fails or when new information contradicting the alignment is acquired. In the Alignment Repair Game (ARG) this evolution is achieved via adaptation operators. ARG was evaluated experimentally and the experiments showed that agents converge towards successful communication and improve their alignments. However, whether the adaptation operators are formally correct, complete or redundant is still an open question. In this paper, we introduce a formal framework based on Dynamic Epistemic Logic that allows us to answer this question. This framework allows us (1) to express the ontologies and alignments used, (2) to model the ARG adaptation operators through announcements and conservative upgrades and (3) to formally establish the correctness, partial redundancy and incompleteness of the adaptation operators in ARG.

Public signature awareness is satisfied if agents are aware of the vocabulary, propositions, used by other agents to think and talk about the world. However, assuming that agents are fully aware of each other's signatures prevents them to adapt their vocabularies to newly gained information, from the environment or learned through agent communication. Therefore this is not realistic for open multi-agent systems. We propose a novel way to model awareness with partial valuations that drops public signature awareness and can model agent signature unawareness, and we give a first view on defining the dynamics of raising and forgetting awareness on this framework.

Partial Dynamic Epistemic Logic allows agents to have different knowledge representations about the world through agent awareness. Agents use their own vocabularies to reason and talk about the world and raise their awareness when confronted with new vocabulary. Through raising awareness the vocabularies of agents are extended, suggesting there is a dual, inverse operator for forgetting awareness that decreases vocabularies. In this paper, we discuss such an operator. Unlike raising awareness, this operator may induce an abstraction on models that removes evidence while preserving conclusions. This is useful to better understand how agents with different knowledge representations communicate with each other, as they may forget the justifications that led them to their conclusions.