Paper deep dive
Holonic Active Distillation for Scalable Multi-Agent Learning in Multi-Sensor Systems
Dani Manjah, Tim Bary, BenoĂŽt Macq, StĂŠphane Galland
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 7/5/2026, 1:34:56 AM
Summary
The paper proposes 'Holonic Active Distillation' (HAD), an architecture within a Holonic Multi-Agent System (HMAS) designed for scalable and adaptive learning in multi-sensor networks. It integrates Clustered Stream-Based Active Distillation (CSBAD) to balance local specialization with global generalization. The framework uses a hierarchical structure of 'holons' (agents) that can self-organize through horizontal clustering (based on sensor stream similarity) and vertical integration (teacher-student relationships). This approach allows the system to handle dynamic sensor additions/departures and manage the trade-off between model specificity and diversity through a multi-tiered architecture.
Entities (7)
Relation Signals (5)
Holonic Active Distillation (HAD) â integrates â Clustered Stream-Based Active Distillation (CSBAD)
confidence 100% ¡ Our approach integrates Clustered Stream-Based Active Distillation (CSBAD)
Holonic Active Distillation (HAD) â ispartof â Holonic Multi-Agent System (HMAS)
confidence 100% ¡ we propose a Holonic Active Distillation architecture within a Holonic Multi-Agent System (HMAS)
Specialized Student â querieslabelsfrom â Teacher
confidence 100% ¡ specialized student models collect local data, query pseudo-labels from teacher models
HAD â uses â ASPECS
confidence 100% ¡ we adopt the ASPECS methodology [9].
Holon â implements â Specialized Student
confidence 90% ¡ The Specialized Student role involves a component tasked with building expertise...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The rapid expansion of sensor-based networks introduces major challenges in scalability, adaptability, and knowledge transfer, especially in open environments where new subsystems can dynamically join or leave. In this work, we propose a Holonic Active Distillation architecture within a Holonic Multi-Agent System (HMAS) to address these issues. Our approach integrates Clustered Stream-Based Active Distillation (CSBAD), a framework in which specialized student models collect local data, query pseudo-labels from teacher models, and cluster into groups of similar sensors. Results show that the holonic organization balances local specialization with global generalization, while efficiently adapting to sensor departures and re-integrations. We also analyzed trade-offs among incremental model updates, system reorganization, and scalability limits. Our findings highlight the advantages of holonic learning for multi-sensor systems while identifying key challenges related to model drift and long-term adaptation.
Tags
Links
- Source: https://arxiv.org/abs/2606.31578v1
- Canonical: https://arxiv.org/abs/2606.31578v1
Trouble viewing inline? Open PDF directly â
Full Text
50,347 characters extracted from source content.
Expand or collapse full text
Holonic Active Distillation for Scalable Multi-Agent Learning in Multi-Sensor Systems Dani Manjah 1 , Tim Bary 1 , Benoit Macq 1 , and StĂŠphane Galland 2 1 Institute of Information and Communication Technologies, Electronics and Applied Mathematics (ICTEAM), UCLouvain, 1348 Louvain-la-Neuve, Belgium dani.manjah, tim.bary, benoit.macq@uclouvain.be 2 UniversitĂŠ de Technologie de Belfort MontbĂŠliard, UTBM, CIAD UR 7533, F-90010 Belfort cedex, France stephane.galland@utbm.fr Abstract. The rapid expansion of sensor-based networks introduces major challenges in scalability, adaptability, and knowledge transfer, es- pecially in open environments where new subsystems can dynamically join or leave. In this work, we propose a Holonic Active Distillation archi- tecture within a Holonic Multi-Agent System (HMAS) to address these issues. Our approach integrates Clustered Stream-Based Active Distilla- tion (CSBAD), a framework in which specialized student models collect local data, query pseudo-labels from teacher models, and cluster into groups of similar sensors. Results show that the holonic organization balances local specialization with global generalization, while efficiently adapting to sensor depar- tures and re-integrations. We also analyzed trade-offs among incremental model updates, system reorganization, and scalability limits. Our findings highlight the advantages of holonic learning for multi-sensor systems while identifying key challenges related to model drift and long- term adaptation. Keywords: Holonic Multi-Agent Systems¡ Distributed Learning¡ Col- laborative Learning¡ Scalable Model Adaptation 1 Introduction In recent years, sensor systems have evolved from isolated and manageable units to expansive and interconnected networks [43]. This transformation, while en- abling broader coverage, challenges the traditional approach of deploying a sin- gle, universal Deep Neural Network (DNN) across all sensors [35]. The dynamic nature of real-world deployments, characterized by stochastic changes and the continuous addition of new sensors, introduces diverse contexts that demand ever larger training datasets and models [27]. This upscaling not only incurs significant costs, but also accumulates hidden technical debt, complicating the maintenance required to adapt to distribution shifts in sensor data [48, 2]. Traditional methods have relied on centralized, monolithic DNNs that strug- gle to scale with the increasing complexity and diversity of sensor networks. arXiv:2606.31578v1 [cs.MA] 30 Jun 2026 2D. Manjah et al. Training system Generalization Specialization Data or Weights Fig. 1. A large-scale distributed training system where DNN nodes are trained on datasets built from a specificity-diversity trade-off for effective learning. Lower nodes are typically tailored to their task (i.e., analytics on a sensor) and operationally more efficient. Higher nodes, trained over vaster and more diverse data, provide generaliza- tion ability. This system can scale up or down, causing challenges in integration and adaptability. These approaches often lead to inefficiencies and increased maintenance chal- lenges. Recent advances, such as the Holonic Learning (HoL) framework, offer a promising alternative by embracing the agent paradigm to improve scalabil- ity and flexibility [14]. HoL leverages self-nested structures of agents, known as holons, to integrate local and global perspectives, facilitating easier subsystem integration and preventing the propagation of disturbances [28, 10, 46]. This hierarchical learning approach improves the efficiency of data and algorithm handling, particularly for large distributed datasets [12]. A question remains re- lated to the design of a scalable distributed learning framework that supports continuous DNN refinement with minimal refactoring, while allowing each unit to update itself online, self-organize with its peers, and transfer knowledge as the system scales. This work builds upon the HoL framework by introducing strategies for ag- gregation, communication, and commitment between learning holons. The con- tributions are twofold: â We augment HoL with standardized organizations and roles, allowing new- comers to integrate knowing only their role and the associated protocol, inspired by active learning and distillation of knowledge [1, 46, 35]. â We propose holonification mechanisms in which agents cluster horizontally and vertically based on the similarity of their sensor streams, preserving confidentiality and balancing specificity and diversity. This recursive pro- HAD for Scalable Multi-Agent Learning3 cess improves the accuracy and resilience of the system, allowing seamless expansion or contraction without disruption [16, 15]. This further advances the adaptability of multisensor holonic systems [37], materializing continuous learning in dynamic environments. The remainder of this paper is structured as follows. Section 2.2 provides background and current related works. Section 3 presents the high-level formula- tion of the organizations for the Holonic Active Distillation (HAD) architecture and the relationships between super- and subholons. Section 4 introduces a dis- tributed, multi-tiered, self-nested structure for DNNs and the self-organization mechanisms that realize the specificityâdiversity trade-off. Section 5 describes the materials used in the experiment, while Section 6 reports the results on sensor addition and removal, comparing partial reorganizations with complete retraining and measuring knowledge-transfer speed. Finally, Section 7 discusses the insights, limitations, and perspectives, followed by the conclusion in Sec- tion 8. 2 Related Works We set the foundational machine learning frameworks to continuously adapt to incoming data. Next, we discuss the landscape of learning in Multi-Agent System (MAS) motivated by the ability of agent-oriented design to minimize refactor- ing and support isolated updates. The section is then concluded by reviewing organizational methodologies for seamless integration of a new sensor, model, or more generally, a subsystem of sensors and models in MAS. 2.1 Online Test-time Adaptation Adapting machine learning models at test-time is crucial to preserve their per- formance [33, 7, 4]. Stream-Based Active Distillation (SBAD) [36] addresses the challenges of the lack of labeled data at test time by selecting, from sen- sor streams, samples to constitute a training set representative of the sensorâs characteristics, i.e., features space. To alleviate annotation costs, SBAD relies on annotation by other models within the system in a Teacher-Student scheme. Fine-tuning a model per device does not scale well, as it requires maintaining a separate model for each additional sensor [18, 48]. Clustered SBAD [35] diminishes the number of models by grouping sensors using a similarity distance of their features and by training one model per cluster. This also improves the accuracy of models by striking a balance between tailor- ing models to their data streams and generalization capabilities by training on enough diverse samples [57]. CSBADâs limitation is not to retain knowledge for successive re-training of current models, nor provide knowledge transfer mecha- nisms to new sensors. Inspired by human organizations, this paper proposes a design where higher layers of systems build structural knowledge to seamlessly integrate agents and avoid the pitfalls of abrupt failures, environmental changes, or knowledge loss, especially with new or varied data classes [19, 30]. 4D. Manjah et al. 2.2 Learning in Multi-Agent Systems Adaptive MAS networks leverage online learning strategies to dynamically re- spond to environmental changes, highlighting the importance of distributed and collaborative learning [39]. Wolpert and Macready [53] introduce a system utilizing reinforcement learn- ing to align agent actions with collective goals, minimizing human oversight. Agents are organized into âsub-worldsâ for focused collaboration, yet the appli- cation of reinforcement learning in complex scenarios with varied sensors and methods encounters obstacles such as unclear rewards and limited exploration, which hinders the required diversity of learning [53, 44, 56]. Organizational learning considers agents evolving through both personal and collective learning efforts, enhancing agentsâ abilities in MAS through man- agement mediated interactions and task alignment to boost system efficiency [21, 50]. This model emphasizes the role of knowledge sharing in improving work- flows and establishing structural knowledge, crucial for system resilience. Social science research [11] reflects on the applicability of this framework to under- stand the impact of staff turnover on management, analogous to agent dynamics in open MAS. Hierarchical learning [13] uses hierarchical MAS to streamline model training in various geographical locations. By modeling challenges as a hypergraph, the system organizes agents, each with unique skills and knowledge, into a structured multitiered network. This design not only facilitates the decentralized handling of Machine Learning (ML) algorithms and data, but also significantly improves the efficiency and scalability of processing large distributed datasets. In the context of distributed ML, Gupta and Raskar [24] pioneered Feder- ated Learning (FL) to train neural networks in distributed datasets, prioritizing data privacy and computational efficiency. However, FL faces hurdles in com- munication and training reliability. Hierarchical FL addresses these by grouping users to improve FL security and efficiency through group-specific updates [55]. Personalized FL [34] methods aim to produce personalized models for different users or groups of users [22] to keep track of their individualized requirements. Hierarchy has also been instrumental in Fog Learning. Unlike FL, which is based on a star topology of device-server interactions, Fog Learning explicitly consid- ers the network and topology structures among devices and enables intelligent device collaborations through data and parameter offloading [25]. Esmaeili et al. [14] abstract FL and CSBAD with HoL, applying holonic prin- ciples to a collaborative learning framework. In that sense, FL and CSBAD can be seen as a first-order HoL. HoL enhances model cooperation with spe- cific strategies for aggregation, communication, and commitment within holons, facilitating complex yet intuitive collaboration of nodes compared to Fog Learn- ing. In this balance between local autonomy and coordinated decision making, holonic systems are better equipped to tackle challenges such as adaptability, and scalability. HoL does not specify how learning agents should (re)organize, nor how a system can seamlessly expand to new domains or safely unlearn obsolete HAD for Scalable Multi-Agent Learning5 ones; shortcomings that become acute in applications requiring auto-scaling and auto-tuning [14, 52]. 2.3 Organizational Multi-Agent Systems Agent-oriented software engineering addresses the limitations of traditional meth- ods like UML in managing complex, distributed, expanding and self-adaptive systems [1, 51, 37, 10, 2, 32, 52, 23]. Organizational theory from social science inspired software designers who developed methodologies for the development of MAS to break down design complexity via 1) the use of metaphors that are more accessible to software engineers and 2) a focus on high levels of abstraction to enable the integration of new agents, even when they differ significantly from existing ones [29, 47, 20]. In fact, MASs are best viewed as organizational struc- tures of autonomous, proactive agents interacting to achieve shared or individual goals [29]. Many agent-oriented methodologies have been developed last decades, such as, ADELFE [3, 23], ASPECS [9], Gaia [54], INGENIAS [42], PASSI [8], Pro- metheus [41], SODA [40], Tropos [6], MOISE [26]. Each has its own specifici- ties: ADELFE is dedicated to adaptive system and cooperative agents design, ASPECS is dedicated to holonic multi-agent systems, Gaia focuses on static organization and roles, whereas PASSI focuses on agent social aspects thanks to ontology, SODA highlights the notion of environment. MOISE focuses on explicit organisational modellingâdefining roles, groups, missions, and deontic normsâto balance agent autonomy and coordinated behaviour. Given our choice of the holonic learning paradigm, we adopt the ASPECS methodology [9]. 2.4 Conclusion We address scalability limitations of continuous adaptation in machine learning systems by enriching the holonic learning paradigm with organizations and roles. These concepts provide an abstract interaction pattern that improves the archi- tectureâs robustness and flexibility. Furthermore, our TeacherâStudent stream- based distillation scheme supplies pseudo-labels that calibrate online incoming Students, thereby enabling auto-tuning. Finally, self-organization emerges from a specificityâdiversity trade-off among Students, while integration and deletion protocols dynamically scale sensor subsystems. Collectively, these mechanisms yield the first HoL variant that supports self-organization and auto- scaling. 3 Holonic Active Distillation Architecture We seek a design that minimizes refactoring and supports isolated updates, sim- plifying the integration of a new sensor, model, or more generally a subsystem of sensors and models [18, 48]. From the literature, we derived five main recom- mendations to design scalable, multimethod learning systems: 6D. Manjah et al. 1. Establish standardized interaction protocols, aggregation strategies, commit- ment, and communication patterns within components. This facilitates the integration of new units, as they only need to understand their role and com- munication methods within the system, regardless of their operating mode [1, 46]. 2. Render a method, a sensor, or by construction a subsystem as independent and self-contained as possible to limit the complexity between units. This aims to simplify a local update or maintenance [1, 23, 51]. 3. Recursively divide a system into subsystems based on a key criterion. This prevents the propagation of disturbances [46]. Furthermore, integration of a new component requires less communication as it requires only coordination with the upper layers of the system instead of with each subsystem [10]. 4. Place units at certain levels of the hierarchy and provide representations of how other levels can contribute âinformationâ or âmodelsâ. This division simplifies the complexity of programming, allowing designers to focus on each module and facilitating reuse between different systems [10]. 5. Exploit Active Distillation, where each Student unit collects data on the fly from its streams to train on them. Training is performed by querying a model Teacher [36]. Given our choice of the organizational holonic paradigm, we adopt the AS- PECS methodology [9]. The latter starts by defining an Organization, which denotes a subsystem in which components play a role and interact to achieve a shared goal in the context of this organization. Next, the Roles which are both expected behaviors to fulfill (part of) requirements, and status to the roleâs agent in the organization (Section 3.1). The subsequent activity (Section 3.2) is the definition of relationships between superholons (higher-level entities) and subholons (lower-level entities). As a reference later, a holarchy denotes the hierarchy of self-regulating holons. 3.1 Teacher-Specialized Student Building on the Active Distillation framework and the specificity-diversity trade- off from [35], we developed an organizational model that incorporates the roles of Specialized Student and Teacher, as illustrated in Fig. 2. The Specialized Student role is designed to continuously collect data on subparts of the systemâs deployment environment. Under the oversight of a higher-order Teacher entity, these Students learn from these data, adapting their modelsâ weights accordingly. 3.2 Holarchy The section begins with introducing a new notation. Then we present an exam- ple of a three-tiered holarchy structure. Each level of this holarchy is a possible instance of an organization defined in Section 3.1. To provide a more holistic HAD for Scalable Multi-Agent Learning7 Active Distillation <<role>> Specialized Student 1,...,N Supervize level n Provide Local Knowledge <<role>> Teacher level n+1 1,...,N Elicit Fig. 2. Organizational model of the Teacher-Student, using the ASPECS notation [9]. The Specialized Student role involves a component tasked with building expertise over a delineated sub-domain in the system, i.e., a regional distribution. The Teacher role supervizes the learning processes of the Students. perspective, we depict the Cyber-Physical Platform (CPP) data processing or- ganization (see our previous work [37]) alongside the TSS but CPP is not the main focus of this paper 3 . Notations A holarchy H L O includes up to L vertical layers instantiating an organization O. A holon i in layer l, where l ranges from 0 to L, is denoted by Ěh l i and comprises: â X l i : Set of operating data streams of a holon Ěh l i . â T l i : Training set of a holon Ěh l i . â V l i : Validation set of a holon Ěh l i . â θ Ěh l i : Processing model of a holon Ěh l i . â SUB l i : Inner members corresponding to layer lâ 1 of a holon Ěh l i . â SUP l i : Superior holon of a holon Ěh l i . Multi-Scale Hierarchical Architecture The system architecture, shown in Fig. 3, includes two holarchies: H 3 CPP that processes data on three levels and H 2 TSS managing knowledge on two levels. â At level 0: agents are the primary functional layer. They employ models designed for specific data streams. Proximity to other agents, geographically or related to the task, allows them to merge outputs and reduce errors. For example, Ěh 0 1 and Ěh 0 2 form Group G1 to fuse their outputs to feed the data request of a higher-order holon Ěh 1 1 . 3 As a more detailed context, CPP is designed to respond to external requests with perceptions and to manage its finite resources to ensure fair access across multiple surveillance operations. The Resource Provider role ensures a fair distribution of the resources among all parties. The Observer role has the ability to produce perceptions thanks to the data acquired by the Sensor role. The data acquisition could be based on another CPP. 8D. Manjah et al. Student Student KNOWLEDGE SPECIALIZA TION DATA SEMANTIC LEVEL G1: Data Fusion Sensor Observer Sensor Observer G2: Data Enrichment G3: Data Enrichment G5: Hierarchical & Distributed Learning Sensor Student Resource Provider GENERALIST SPECIFIC HOLISTIC LOCAL G4: Knowledge Distillation Teacher Fig. 3. Holonic architecture inspired by the âcheese boardâ notation [9, 17]. Each level represents a different hierarchical position, defining both the semantic level of data and the degree of knowledge specialization. On the left, the H 3 CPP instantiates the CPP organization, and on the right, the H 2 TS is responsible for active learning. Agents may assume multiple roles and participate in multiple holarchies simultaneously. However, agents monitoring the same area may employ different models if their functions require learning different features. Consequently, Ěh 0 2 and Ěh 0 3 , as Specialized Students, form Group G4 to learn a shared model under the supervision of a Teacher holon via Group G5. â At levels 1 and 2: higher levels above the agents integrate and synthe- size data from specific areas of the system (e.g., data streams that share attributes). The holons in the role Observer, such as Ěh 1 1 and Ěh 2 1 , elevate the collected data to a new semantic level. Holon Ěh 1 2 , a higher order Specialized Student, aggregates validation sets from Ěh 0 2 and Ěh 0 3 (i.e., T 1 2 = V 0 2 âŞV 0 3 ), to create a broader and generalized model. Generally, each semantic level consolidates knowledge across broader areas of the system, fostering a holistic view, such as city-scale tracking. Meanwhile, intermediate layers consisting of Specialized Students synthesize knowledge from lower levels, to deepen collective task understanding and increase holonsâ universality. 4 Holonic Learning Framework This section presents mechanisms to create a multilevel learning framework. Next, it introduces a mechanism to incorporate new nodes by coordinating with the top layers and assigning each new node to the group whose DNN model is most accurate in its data stream. HAD for Scalable Multi-Agent Learning9 4.1 Holonification In the holonic terminology, holonification is the process of grouping agents into a holarchy, resembling complex clustering based on criteria like capabilities and resource access [16, 15]. In this Section, we propose a multi-tiered learning structure (illustrated in Fig. 1), comprising a portfolio of models that range from sensor-specific to uni- versal, deployable across the entire network. Specifically, upper-layer models are trained on larger datasets for broader coverage, while lower-layer models use smaller, more similar datasets for increased specificity. Having intermediate models at various levels of granularity not only ensures adaptability, but also supports robust knowledge organization. For example, city-wide vehicle detection may require multiple models specializing in certain domain represen- tations [35]. However, these domain-specific models benefit from interactions with peer models or a more fundamental model that develops a funda- mental understanding of object detection tasks [57]. Our agent-based modeling offers this flexibility to develop these vertical and horizontal interactions. Formally, each holon in a layer l > 0 is allocated a budget B l = B 0 ¡ 10 l , where B 0 represents the number of images used for model fine-tuning. This budget limits the training of each layer to at most 10 l from the preceding levels, ensuring that the size of the data set of any holon Ěh l i does not exceed B l , i.e., |T l i |⤠B l . To merge holons, we adopt the premise from Manjah et al. [35] that models with similar performance have learned from comparable data. The remainder of this section describes the holonification process. STEP 1 â Cross-Performance Vector. Assuming holons can transfer their model weights to each other within the same layer. Each holon Ěh l i computes a perfor- mance vector P i by evaluating the effectiveness of models from other holons and itself in the same layer on its own validation data V l i , according to Equation 1. P l i := f θ l 1 ,V l i ;¡ ;f θ l N l ,V l i T (1) Where θ j denotes the model parameters of the j-th holon Ěh l j , j â 1,¡ ,N l and f (θ,V) the score of a model performance θ against a validation set V. STEP 2 â Pair-Wise Distance Computation. To quantify the differences be- tween models trained in different domains, holons broadcast their cross-perfor- mance vectors P l i defined in Equation 1 and compute a pairwise distance between their performance in the datasets and the performance of the other holons. Gen- erally, for a holon Ěh l i , the distance to a holon Ěh l j is given by Equation 2. D i ( Ěh l j ) = v u u t N l X k=1 (P l ik â P l jk ) 2 (2) 10D. Manjah et al. STEP 3 â Agglomerative Merging using Single Linkage. The merging of the holons is an iterative process. The set of holons Ěh l 1 ,¡ , Ěh l N l creates a higher- order holarchy H l+1 to which they belong. At each iteration, the set of holons Ěh l transmits their smallest linkage dis- tance. This is defined as the minimal distance between the inner members of the holons. Formally, for two holons Ěh l A and Ěh l B , the single link distance L( Ěh l A , Ěh l B ) is given by Equation 3. L( Ěh l A , Ěh l B ) = minD ij : Ěh lâ1 i â SUB( Ěh l A ), Ěh lâ1 j â SUB( Ěh l B )(3) After all linkages are evaluated, the pair with the smallest distance merges, involving a combination of their datasets. After merging, the set of holons has decreased, Ěh l 1 ,¡ , Ěh l N l â1 , and the linkage distances are updated for all agents. The process ends if there remains only one holon or if the previous merge leads to a holon with a dataset size that exceeds B l+1 . In the second scenario, the process goes back to STEP 1 for the set Ěh l+1 1 ,¡ , Ěh l N l+1 . STEP 4 â Model Training. The final steps consist in training the cluster models on the aggregated data sets. 4.2 Domain Integration Process A new holon Ěh + joins a holarchy H L of L levels. Its integration starts at the highest hierarchical level, L, and progresses downward to the level 1. At each level, Ěh + is associated with the holon Ěh 1 â that shows the highest performance in the new set of unit validations, V + , subject to meeting budget constraints, i.e., |T l â âŞT + | ⤠B l . Once integrated, Ěh + âs dataset merges with that of the selected holon, Ěh l â , necessitating a retraining of the aggregated dataset. If no appropriate holons are available at a required level, the system can initiate reholonification, integrating Ěh + with the set Ěh l 1 ,¡ , Ěh l N l . A pseudocode is provided in Algo- rithm 1. Remark 1. The cost of holonification is compared to integration on-the-fly on the basis of the amount of communication between the holons. It is built on a single linkage-Hierarchical Clustering, with a complexity of O N 2 [49]. On the other hand, the on-the-fly mechanism has a O(N + L) com- plexity. This corresponds to the worst-case scenario in which the new agent is compared to all holons from the upper layer L to layer 0. This mechanism thus offers a cost-effective integration in comparison with a reholonification. 4.3 Research Questions From the setup and challenges described above, we formulate the following re- search questions: 1. Given a new data stream, how can we determine the most suitable existing model for fine-tuning? HAD for Scalable Multi-Agent Learning11 Algorithm 1 Integration of a New Holon into a Holarchy Require: H L : L-level learning holarchy Require: Ěh + : A holon 1: for l = L downto 1 do ⡠Identify sub-holons whose training sets do not exceed budget constraints 2:FreeHolonsâ Ěh l :|T l âŞT + |⤠B l 3: if FreeHolons =â then 4:Reholonification with the set Ěh l 1 ,¡ , Ěh l N l ⪠Ěh + . 5:break 6: end if ⡠Select the optimal sub-holon for integration 7:FreeModelsâ FreeHolonsâs models 8: θ l â â arg max θ l âFreeModels f θ l ,V + 9: Update model parameters θ l â using T l â âŞT + . ⡠Integrate Ěh + into holon of θ l â 10:SUB( Ěh l â )â SUB( Ěh l â )⪠Ěh + 11: end for 2. Assume an effective integration of new sensors based on similarity with a group of sensors from the system: (a) What are the consequences on model accuracy upon the integration of a new agent? (b) How does the accuracy of the model scale when incrementally integrating N + new agents versus performing a full system reorganization? 3. What are the long-term accuracy trade-offs between retaining versus dis- carding data from removed sensors? 5 Materials and Methods The datasets, the training procedure, and the evaluation protocol are presented in this Section. 5.1 Datasets We used two city-focused video datasets for a total of 16 cameras. WALT [45] features footage from nine static cameras over 1â4 weeks. Sampling rates vary (5,000â40,000 frames/week), with temporal bursts and diverse weather conditions (snow, rain, day/night). AI-City [38] features seven annotated videos, each approximately five minutes at 10 FPS. Camera angles and sensor types vary (vertical, dome, PTZ), ensuring coverage of multiple representation contexts. 12D. Manjah et al. 5.2 Model Training We follow SBAD [35] sampling 256 images per camera. A large YOLOv8x6 Teacher (261.1 GFLOPs) pseudolabels these samples. Each Student model is a YOLOv8n (8.7 GFLOPs), initialized with COCO weights [31], then fine-tuned at a learning rate of 0.01 (unless otherwise indicated). 5.3 Evaluation of mAP50-95 We report the âmAP50-95" as the mean Average Precision across various in- tersection over union thresholds, spanning from 0.50 to 0.95 in increments of 0.05 . We evaluated the holonâs performance on its associated datasets. 6 Results We begin by evaluating our holonification approach under different budgets and then proceed with incremental integration, departure handling, and knowledge- transfer experiments. 6.1 Holonification Baseline Performance We conducted a holonification, as proposed in Section 4.1, on a dataset com- prising sixteen cameras. We set a multilayer budget framework B l = 256¡ 10 l , implying that layer 0 holons do not exceed 256 training samples, and successive layers cannot exceed 10 l sub-holons for a holon Ěh l . Table 1 confirms that the Table 1. Holonification with varying budgets. Shown are the final groupings and av- erage mAP50-95 for the 16-camera dataset. LayerB l mAP50-95 Holonic Structure 2256000.65 H 2 : Ěh 1 0 , Ěh 1 1 , Ěh 1 2 125600.66 H 1 0 : Ěh 0 0 , Ěh 0 1 , Ěh 0 2 , Ěh 0 8 H 1 1 : Ěh 0 3 , Ěh 0 4 , Ěh 0 5 , Ěh 0 6 , Ěh 0 7 H 1 2 : Ěh 0 9 , Ěh 0 10 , Ěh 0 11 , Ěh 0 12 , Ěh 0 13 , Ěh 0 14 , Ěh 0 15 02560.67 H 1 : Ěh 0 0 ,..., Ěh 0 15 YOLOv8n COCO N.A0.498 N.A models require specificity to achieve maximum performance. HAD for Scalable Multi-Agent Learning13 6.2 Transfer, Integration and Departure Model Transfer Upon Increment The transferability of holons across new, although similar, domains is investigated. Table 2 details the performance results for models trained in an all-but-one combination of domains as well as across all domains. Table 2. mAP50-95 scores for models trained under an all-but-one camera to assess the transferability of those models on the remaining camera. A baseline is also pro- vided where the model is trained across all cameras. Each model is trained for 10000 iterations. f (θ,V i ) ClusterV 0 0 V 0 1 V 0 2 V 0 3 θ H 1 1 \ Ěh 0 0 0.38 0.65 0.65 0.47 θ H 1 1 \ Ěh 0 1 0.46 0.57 0.66 0.46 θ H 1 1 \ Ěh 0 2 0.49 0.66 0.65 0.48 θ H 1 1 \ Ěh 0 8 0.47 0.64 0.66 0.42 θ H 1 1 0.46 0.65 0.66 0.47 The results indicate that models struggle to transfer, even across similar camera domains, reinforcing the need to integrate the newcomer in a cluster, and the local retraining the cluster. Incremental Integration We evaluated the impact of integrating N + new units into a holonified system, structured with budget limits of B l = 256¡ 10 l . Using our integration mechanism described in Algorithm 1, we evaluated two scenarios: integrating one (N + = 1) and three (N + = 3) additional agents. In 16 agent configurations, the incremental integration maintained an average mAP50-95 of 0.66Âą 0.003 (N + = 1) and 0.66Âą 0.006 (N + = 3), showing no degradation compared to the baseline in Table 1. Agent Departure When a sensor Ěh 0 i leaves, its data T i may be retained or discarded. We successively simulate the exit of each agent and track the accuracy of the global model on (i) remaining and (i) left sensors. As Fig. 4 shows, removing a sensorâs data yields small gains for the remaining sensors, but severely reduces performance if that sensor later re-enters the system. Note that, upon the departure of an agent, their data set T 0 i is removed from the collective data set T 2 , and the model is re-trained for 10,000 iterations at a learning rate of 0.005. 14D. Manjah et al. 0123456789101112131415 Number of departed holons 0.25 0.20 0.15 0.10 0.05 0.00 0.05 Difference in mAP50-95 Average Difference in mAP50-95 Between both Retraining Alternatives on Test Sets of Departed and Non-Departed Holons Non-departed holons Departed holons Fig. 4. Difference in Ěh 2 model performance between retaining and discarding each departed sensorâs data. Blue: remaining sensors; yellow: departed sensors. Results show marginal gains for remaining sensors but a marked degradation on departed sensors. 6.3 Inter-Holonic Knowledge Transfer We test how effectively a holon trained on existing cameras can accelerate train- ing and improve the peak accuracy of a newcomer domain. Specifically, we con- ducted 16 trials, each excluding one camera from the data set to simulate a ânewcomerâ. The following pre-trained models serve as initial weights: â θ 2 : Global holon (trained on 15 cameras), â θ 1 â : Group-specific holon, â YOLOv8n COCO : General-purpose off-the-shelf model. Fig. 5 shows that θ 2 or θ 1 â consistently outperform the generic COCO base- line when fine-tuning the newcomer camera. Training spanned 5 epochs with a learning rate of 0.005. 7 Discussion 7.1 Insights Our experiments confirm that a certain level of domain specificity improves ac- curacy (Section 6.1), though it also increases the number of models to maintain. Budget constraints help contain this growth but can reduce performance gains from specialized holons. Meanwhile, leveraging broader universal models accel- erates learning for new domains (Section 6.3). In the context of open systems (Section 6.2), our sanity check shows that a straightforward model transfer performs under, even when the model comes HAD for Scalable Multi-Agent Learning15 012345 Epoch 0.1 0.2 0.3 0.4 0.5 0.6 mAP50-95 mAP50-95 per Epoch for Different Pretrained Weights Pre-Trained Weights 2 1 * COCO Fig. 5. mAP50-95 per epoch for a new model starting from universal model θ 2 , group- specific θ 1 â , and general-purpose θ COCO . The superiority of θ 2 highlights the efficiency of selecting a pretrained model closer to the source. from similar domains. The observed performance gap motivated the development of integration mechanisms, which proved effective, but the experiments do not provide conclusive evidence regarding the maximum number of agents that can be integrated without a performance decline. Finally, agent departure highlights a trade-off between short-term gains and relearning costs if the environment reappears. In other words, discarding data should be considered in terms of agentsâ turnover rate. 7.2 Limitations Machine Learning Lifecycle Machine learning based models are also subject to feedback loops, where data and interactions with the external world influence their behavior in unintentional ways [48]. The subsequent design of machine learning systems should account for the fact that their behavior evolves with environmental data and user interactions. This includes providing control mech- anisms to avoid the accumulation of errors due to the self-supervised nature of the system. Stress-tests We need to further stress test the system; that is, starting with a system of size N, stress tests can evaluate how many new components (N + ) can be integrated without compromising the quality of service. 7.3 Perspectives Modern systems integrate heterogeneous approaches (e.g., physics-based model- ing vs. deep learning) and diverse sensing modalities (e.g., cameras, radar), pro- 16D. Manjah et al. ducing richer analytics [5, 37]. Our architecture abstracts the holonic paradigm sufficiently to accommodate such heterogeneity. However, specialized coordina- tion modules could further optimize collaborative performance among different modalities. 8 Conclusions We present an organizational holonic learning design coupled with active learn- ing to address the challenges of scaling learning in multisensor networks. Our self-organization mechanism, grounded in the specificity-diversity trade-off, al- lows for the establishment of various granularity levels and handles sensor addi- tion and removal while maintaining strong predictive performance. Experimental results highlight the benefits of vertical and horizontal knowledge transfer, al- though more stress testing is needed to refine the upper limits on system growth. We also note that self-supervised processes risk model drift without robust mon- itoring, which may cause issues in autoscaling and autotuning. Future work aims to design colearning mechanisms for heterogeneous methods. Acknowledgments. This work was partially funded by PIT ATMP - Convention 8881. T. Bary is funded by the MedReSyst project, supported by FEDER and the Walloon Region. Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article. Bibliography [1] Abbas, H.A.: Organization of multi-agent systems: An overview. Interna- tional Journal of Intelligent Information Systems 4(3), 46 (2015), https: //doi.org/10.11648/j.ijiis.20150403.11 [2] Beal, J., Viroli, M., Pianini, D., Damiani, F.: Self-adaptation to device dis- tribution changes. In: 2016 IEEE 10th International Conference on Self- Adaptive and Self-Organizing Systems (SASO), p. 60â69 (2016), https: //doi.org/10.1109/SASO.2016.12 [3] Bernon, C., Gleizes, M.P., Peyruqueou, S., Picard, G.: Adelfe: A methodol- ogy for adaptive multi-agent systems engineering. In: Petta, P., Tolksdorf, R., Zambonelli, F. (eds.) Engineering Societies in the Agents World I, p. 156â169, Springer Berlin Heidelberg, Berlin, Heidelberg (2003), ISBN 978-3-540-39173-9, https://doi.org/10.1007/3-540-39173-8_12 [4] Brion, E., LĂŠger, J., Javaid, U., Lee, J., Vleeschouwer, C.D., Macq, B.: Using planning CTs to enhance CNN-based bladder segmentation on cone beam CT. In: Fei, B., Linte, C.A. (eds.) Medical Imaging 2019: Image-Guided Procedures, Robotic Interventions, and Modeling, vol. 10951, p. 109511M, International Society for Optics and Photonics, SPIE (2019), https://doi. org/10.1117/12.2512791, URL https://doi.org/10.1117/12.2512791 [5] Campagner, A., Ciucci, D., Cabitza, F.: Aggregation models in ensem- ble learning: A large-scale comparison. Information Fusion 90, 241â252 (2023), ISSN 1566-2535, https://doi.org/https://doi.org/10.1016/j. inffus.2022.09.015 [6] Castro, J., Kolp, M., Mylopoulos, J.: Towards requirements-driven in- formation systems engineering: the tropos project. Information Sys- tems 27(6), 365â389 (2002), ISSN 0306-4379, https://doi.org/10.1016/ S0306-4379(02)00012-1 [7] Cioppa, A., Deliege, A., Istasse, M., De Vleeschouwer, C., Van Droogen- broeck, M.: Arthus: Adaptive real-time human segmentation in sports through online distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops (June 2019) [8] Cossentino, M., Gaglio, S., Sabatucci, L., Seidita, V.: The passi and ag- ile passi mas meta-models compared with a unifying proposal. In: Multi- Agent Systems and Applications IV, p. 183â192, Springer Berlin Heidel- berg, Berlin, Heidelberg (2005), ISBN 978-3-540-31731-9 [9] Cossentino, M., Gaud, N., Hilaire, V., Galland, S., Koukam, A.: Aspecs: an agent-oriented software process for engineering complex systems. Au- tonomous Agents and Multi-Agent Systems 20(2), 260â304 (2010), https: //doi.org/10.1007/s10458-009-9099-4 [10] Diaconescu, A., Frey, S., MĂźller-Schloer, C., Pitt, J., Tomforde, S.: Goal- oriented holonics for complex system (self-)integration: Concepts and case studies. In: 2016 IEEE 10th International Conference on Self-Adaptive and 18D. Manjah et al. Self-Organizing Systems (SASO), p. 100â109 (2016), https://doi.org/ 10.1109/SASO.2016.16 [11] Dong, J., Liu, R., Qiu, Y., Crossan, M.: Should knowledge be distorted? managersâ knowledge distortion strategies and organizational learning in different environments. The Leadership Quarterly 32(3), 101477 (2021), ISSN 1048-9843 [12] Esmaeili, A., Gallagher, J.C., Springer, J.A., Matson, E.T.: Hamlet: A hier- archical agent-based machine learning platform. ACM Trans. Auton. Adapt. Syst. 16(3â4) (jul 2022), ISSN 1556-4665 [13] Esmaeili, A., Gallagher, J.C., Springer, J.A., Matson, E.T.: Hamlet: A hierarchical agent-based machine learning platform. ACM Trans. Auton. Adapt. Syst. 16(3â4) (Jul 2022), ISSN 1556-4665, https://doi.org/10. 1145/3530191 [14] Esmaeili, A., Ghorrati, Z., Matson, E.T.: Holonic learning: A flexible agent- based distributed machine learning framework. In: Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems, p. 525â533, AAMAS â24, International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC (2024), ISBN 9798400704864 [15] Esmaeili, A., Mozayani, N., Jahed-Motlagh, M.R., Matson, E.T.: Towards topological analysis of networked holonic multi-agent systems. In: Ad- vances in Practical Applications of Survivable Agents and Multi-Agent Sys- tems: The PAAMS Collection, p. 42â54, Springer International Publishing, Cham (2019), ISBN 978-3-030-24209-1 [16] Esmaeili, A., Mozayani, N., Motlagh, M.R.J., Matson, E.T.: The impact of diversity on performance of holonic multi-agent systems. Engineering Applications of Artificial Intelligence 55, 186â201 (2016), ISSN 0952-1976 [17] Feraud, M., Galland, S.: First comparison of sarl to other agent- programming languages and frameworks. Procedia Computer Science 109, 1080 â 1085 (2017), ISSN 1877-0509, https://doi.org/https://doi.org/ 10.1016/j.procs.2017.05.389 [18] Fowler, M.: Refactoring: improving the design of existing code. Addison- Wesley Professional (2018) [19] French, R.M.: Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences 3(4), 128â135 (1999), ISSN 1364-6613, https://doi. org/10.1016/S1364-6613(99)01294-2 [20] Garcia, E., Argente, E., Giret, A., Botti, V.: Issues for organizational multi- agent systems development. In: Sixth International Workshop From Agent Theory to Agent Implementation (AT2AI-6), p. 59â65, Citeseer (2008) [21] Gherardi, S.: Learning: Organizational. In: International Encyclopedia of the Social & Behavioral Sciences (Second Edition), p. 695â698, Elsevier, Oxford, second edition edn. (2015), ISBN 978-0-08-097087-5 [22] Ghosh, A., Chung, J., Yin, D., Ramchandran, K.: An efficient framework for clustered federated learning. In: Advances in Neural Information Processing Systems, vol. 33, p. 19586â19597, Curran Associates, Inc. (2020) [23] Gleizes, M.P.: Self-adaptive complex systems. In: Multi-Agent Systems, p. 114â128, Springer, Berlin, Heidelberg (2012), ISBN 978-3-642-34799-3 HAD for Scalable Multi-Agent Learning19 [24] Gupta, O., Raskar, R.: Distributed learning of deep neural network over multiple agents. Journal of Network and Computer Applications 116, 1â8 (2018), ISSN 1084-8045, https://doi.org/10.1016/j.jnca.2018.05.003 [25] Hosseinalipour, S., Brinton, C.G., Aggarwal, V., Dai, H., Chiang, M.: From federated to fog learning: Distributed machine learning over heterogeneous wireless networks. IEEE Communications Magazine 58(12), 41â47 (2020) [26] HĂźbner, J.F., Boissier, O., Kitio, R., Ricci, A.: Instrumenting multi-agent organisations with organisational artifacts and agents: âgiving the organ- isational power back to the agentsâ. Autonomous agents and multi-agent systems 20(3), 369â400 (2010) [27] Jocher, G., Chaurasia, A., Qiu, J.: Ultralytics YOLO (2023), URL https: //github.com/ultralytics/ultralytics [28] Koestler, A.: The ghost in the machine. Hutchinson, London, UK (1967) [29] Kolp, M., Giorgini, P., Mylopoulos, J.: Multi-agent architectures as organi- zational structures. Autonomous Agents and Multi-Agent Systems 13, 3â25 (2006) [30] Le, J., Lei, X., Mu, N., Zhang, H., Zeng, K., Liao, X.: Federated continuous learning with broad network architecture. IEEE Transactions on Cybernet- ics 51(8), 3874â3888 (2021), ISSN 2168-2275, https://doi.org/10.1109/ TCYB.2021.3090260 [31] Lin, T., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P., Ramanan, D., DollĂĄr, P., Zitnick, C.L.: Microsoft COCO: common objects in context. CoRR abs/1405.0312 (2014) [32] Lippi, M., Mariani, S., Martinelli, M., Zambonelli, F.: Individual and col- lective self-development: Concepts and challenges. In: 2022 17th Conference on Computer Science and Intelligence Systems (FedCSIS), p. 15â21 (2022) [33] LĂŠger, J., Brion, E., Desbordes, P., De Vleeschouwer, C., Lee, J.A., Macq, B.: Cross-domain data augmentation for deep-learning-based male pelvic organ segmentation in cone beam ct. Applied Sciences 10(3) (2020), ISSN 2076-3417, https://doi.org/10.3390/app10031154, URL https://w. mdpi.com/2076-3417/10/3/1154 [34] Ma, Z., Xu, Y., Xu, H., Liu, J., Xue, Y.: Like attracts like: Personalized federated learning in decentralized edge computing. IEEE Transactions on Mobile Computing 23(2), 1080â1096 (2024), ISSN 1558-0660 [35] Manjah, D., Cacciarelli, D., De Vleeschouwer, C., Macq, B.: Camera cluster- ing for scalable stream-based active distillation. Expert Systems with Ap- plications 290, 128408 (2025), ISSN 0957-4174, https://doi.org/https: //doi.org/10.1016/j.eswa.2025.128408 [36] Manjah, D., Cacciarelli, D., Standaert, B., Benkedadra, M., de Hertaing, G.R., Macq, B., Galland, S., De Vleeschouwer, C.: Stream-based active dis- tillation for scalable model deployment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Work- shops, p. 4998â5006 (2023) [37] Manjah, D., Galland, S., Vleeschouwer, C.D., Macq, B.: Autonomous meth- ods in multisensor architecture for smart surveillance. In: Proceedings of the 16th International Conference on Agents and Artificial Intelligence, vol. 3, p. 824â832 (2024), ISBN 978-989-758-680-4, ISSN 2184-433X 20D. Manjah et al. [38] Naphade, M., Wang, S., Anastasiu, D.C., Tang, Z., Chang, M.C., Yang, X., Yao, Y., Zheng, L., Chakraborty, P., Lopez, C.E., Sharma, A., Feng, Q., Ablavsky, V., Sclaroff, S.: The 5th ai city challenge. In: The IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR) Workshops (2021) [39] Nezamoddini, N., Gholami, A.: A survey of adaptive multi-agent networks and their applications in smart cities. Smart Cities 5(1), 318â347 (2022), ISSN 2624-6511 [40] Omicini, A.: Soda: Societies and infrastructures in the analysis and design of agent-based systems. In: Agent-Oriented Software Engineering, p. 185â 193, Springer Berlin Heidelberg (2001), ISBN 978-3-540-44564-7 [41] Padgham, L., Winikoff, M.: Prometheus: A methodology for developing intelligent agents. In: Agent-Oriented Software Engineering I, p. 174â 185, Springer Berlin Heidelberg, Berlin, Heidelberg (2003), ISBN 978-3- 540-36540-2 [42] PavĂłn, J., GĂłmez-Sanz, J.: Agent oriented software engineering with inge- nias. In: Multi-Agent Systems and Applications I, p. 394â403, Springer, Berlin, Heidelberg (2003), ISBN 978-3-540-45023-8 [43] Perera, C., Zaslavsky, A., Christen, P., Georgakopoulos, D.: Sensing as a service model for smart cities supported by Internet of Things. Transactions on Emerging Telecommunications Technologies 25(1), 81â93 (2014), ISSN 2161-3915 [44] Porter, B., Rodrigues Filho, R.: Distributed emergent software: Assembling, perceiving and learning systems at scale. In: 2019 IEEE 13th International Conference on Self-Adaptive and Self-Organizing Systems (SASO), p. 127â 136 (2019) [45] Reddy, N.D., Tamburo, R., Narasimhan, S.G.: Walt: Watch and learn 2d amodal representation from time-lapse imagery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 9356â9366 (2022) [46] Rodriguez, S., Hilaire, V., Gaud, N., Galland, S., Koukam, A.: Holonic Multi-Agent Systems. Natural Computing Series 37, 251â279 (2011), https://doi.org/10.1007/978-3-642-17348-6_11 [47] Schatten, M., Grd, P., Konecki, M., KudeliÄ, R.: Towards a formal conceptu- alization of organizational design techniques for large scale multi agent sys- tems. Procedia Technology 15, 576â585 (2014), ISSN 2212-0173, https:// doi.org/10.1016/j.protcy.2014.09.018, 2nd International Conference on System-Integrated Intelligence: Challenges for Product and Production Engineering [48] Sculley, D., Holt, G., Golovin, D., Davydov, E., Phillips, T., Ebner, D., Chaudhary, V., Young, M., Crespo, J.F., Dennison, D.: Hidden technical debt in machine learning systems. In: Advances in Neural Information Pro- cessing Systems, vol. 28, Curran Associates, Inc. (2015) [49] Sibson, R.: SLINK: An optimally efficient algorithm for the single-link clus- ter method. The Computer Journal 16(1), 30â34 (1973), ISSN 0010-4620 HAD for Scalable Multi-Agent Learning21 [50] Terabe, M., Washio, T., Katai, O., Sawaragi, T.: A study of organiza- tional learning in multiagents systems. In: Distributed Artificial Intelligence Meets Machine Learning Learning in Multi-Agent Environments, p. 168â 179, Springer Berlin Heidelberg, Berlin, Heidelberg (1997), ISBN 978-3-540- 69050-4 [51] Wautelet, Y., Schinckus, C., Kolp, M.: Agent-based software engineering, paradigm shift, or research program evolution. In: Research Anthology on Recent Trends, Tools, and Implications of Computer Programming, p. 1642â1654, IGI Global (2021) [52] Weyns, D., Gerostathopoulos, I., Abbas, N., Andersson, J., Biffl, S., Brada, P., Bures, T., Di Salle, A., Galster, M., Lago, P., Lewis, G., Litoiu, M., Musil, A., Musil, J., Patros, P., Pelliccione, P.: Self-adaptation in industry: A survey. ACM Trans. Auton. Adapt. Syst. 18(2) (2023), ISSN 1556-4665 [53] Wolpert, D.H., Macready, W.G.: No free lunch theorems for optimization. IEEE Transactions on Evolutionary Computation 1(1), 67â82 (1997), ISSN 1089778X, https://doi.org/10.1109/4235.585893 [54] Wooldridge, M., Jennings, N.R., Kinny, D.: The gaia methodology for agent- oriented analysis and design. Autonomous Agents and Multi-Agent Systems 3(3), 285â312 (2000), https://doi.org/10.1023/A:1010071910869 [55] Xu, B., Xia, W., Wen, W., Liu, P., Zhao, H., Zhu, H.: Adaptive hierarchical federated learning over wireless networks. IEEE Transactions on Vehicular Technology 71(2), 2070â2083 (2021) [56] Yang, Y., Luo, J., Wen, Y., Slumbers, O., Graves, D., Bou Ammar, H., Wang, J., Taylor, M.E.: Diverse auto-curriculum is critical for successful real-world multiagent learning systems. In: Proceedings of the 20th Inter- national Conference on Autonomous Agents and MultiAgent Systems, p. 51â56, AAMAS â21, International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC (2021), ISBN 9781450383073 [57] Yin, D., Pananjady, A., Lam, M., Papailiopoulos, D., Ramchandran, K., Bartlett, P.: Gradient diversity: a key ingredient for scalable distributed learning. In: Proceedings of the Twenty-First International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Re- search, vol. 84, p. 1998â2007, PMLR (2018)