WorksComputer Science
Let's Talk This Out: Towards Explanation-based Distributed Goal Satisfaction
Abstract
Digital systems are embedded in a modular environment of different components with individual goals and settings, which are often installed without central oversight. In this research preview paper we present a framework for communication between components to facilitate compositional problem solving based on explanation patterns and A/G-contracts, without the need for an oracle. Our approach uses partial MaxSMT solving and quantifier elimination to compute counterfactual explanation messages. We illustrate the feasibility of the approach with a proof-of-concept chain topology example, and discuss future challenges when expanding the protocol to more complex topologies.
Introduction
Cyber-physical systems (CPS) exist in, and interact with, their environment and other CPSs in it. Because of these interactions, software verification methods are incomplete that only take a single component of the greater system into account, without considering its effects on and influences from the environment. The state of the environment and its possible natural evolution, as well as the actions and reactions of other components in it, lead to a state space explosion that makes it infeasible to derive the correct set of actions during development. At the same time, the precise composition of the system and all its components is generally unknown beforehand if we consider modular systems where a user may arrange and replace components dynamically.
Take, for example, a modern energy infrastructure utilizing a broad range of decentralized renewable energy sources. The individual generators may come from different producers and editions, but still be connected to the same electricity grid, and supply machines which are often even more diverse in their origins and function. Even if one could plan the whole infrastructure at one point, it is bound to change due to replacements or repairs, or even extreme changes in circumstances such as storms, or unusually long stretches of time without direct sunlight. How can the correct cooperation between all of these components then be guaranteed?
We propose a five-step workflow for the generation and use of explanations based on assume-guarantee contracts (A/G-contracts), minimal correction sets (MCS), and quantifier elimination (QE): (1) a component detects a conflict between its environment assumptions and the target it guarantees; (2) it identifies a responsible MCS candidate among its environment assumptions; (3) it computes maximally permissive relaxations of these assumptions via QE; (4) it sends them as a counterfactual explanation to the component controlling the assumption; and (5) the receiver either returns a new value satisfying the relaxed constraint and its own target, or a rejection.
The explanation generation then needs to be embedded in a communication protocol which terminates and either finds a new solution or proves that none exists. We present the generation of maximally permissive counterfactual explanations between two components, and discuss challenges arising in more complex topologies.
We impose the following rules on the system and interactions between components: First, we assume that the system was in a working state where every target was satisfied before a disturbance introduced a fault that now needs to be compensated. Second, each component knows its own interface: which shared variables affect it, which neighbour controls each of them, an exact model of how these variables affect its own target, and the current values of all quantities that affect it but that it does not control (e.g., a measured wind speed). It knows neither the internal models of other components nor the global topology. Learning approaches or separate communications could be used to determine these interactions.
Related Work
Interactions between components in a system of systems are an active topic of research, which has been growing in recent years. However, most existing research directions treat it as a game-theoretic problem. Of existing work, we found that Chatterjee et al. [7] promote a similar concept as we do, i.e., the gradual weakening of environment assumptions through minimal edge removal on a game graph, which mirrors our use of MCSs. Recently, Mu et al. [21] have explored mechanisms to assign responsibility for outcomes to individual actions of components in a system, which is close in concept to our complaint mechanism. While their approach is post-hoc, Parker et al. [22] focus on anticipatory responsibility attribution. Finally, some works lie at the intersection of game theory and A/G-contracts, for example Cavezza et al. [6]. Anand et al. [2] elaborate on the requirements for synthesized assumptions, i.e., sufficiency, actionability, and permissiveness.
Distributed constraint satisfaction and optimization (DisCSP/DCOP) [28, 12] also coordinates agents that each control some variables of a shared constraint problem. In contrast to that setting, our components pursue heterogeneous local targets and communication is only initiated when one fails to reach it. The only global objective is the fulfillment of all local targets.
A/G-contracts for compositional reasoning were introduced by Misra and Chandy [20] and Abadi and Lamport [1], and actively used for CPS design since Sangiovanni-Vincentelli et al. [24]. Modern meta-theory and different formalisms are presented by Benveniste et al. [5]. Luckcuck et al. [19] combine first-order A/G-contracts, compositional reasoning, and runtime monitoring for modular robotic systems. Like ours, their setting is modular and contract-based. However, their contracts are defined at design time and monitored at runtime, whereas we use violated assumptions as the starting point for negotiating new ones at runtime. Their work also provides valuable insight into different network topologies and the challenges one needs to overcome when considering them.
A/G-contracts need to specify environment assumptions for which the defined properties are guaranteed to hold. Therefore, the learning of assumptions which are consistent with all components in a given system is a topic of specific interest. Giannakopoulou et al. [15] introduced the framework of assumption learning, which was refined by Cobleigh et al. [9] to use Angluin’s algorithm [3]. Several further refinements and applications of this approach exist, such as the extension to probabilistic assumptions by Feng et al. [10], and the extension to timed automata by Chen et al. [8]. In particular, Frenkel et al. [13] introduce a repair step which is invoked when a bad trace is found during verification. Diverging from the -style assumption learning line, Bobaru et al. [14] instead use a CEGAR-style abstraction refinement. However, what all approaches have in common is that they take place during development time and require an “omniscient” oracle which knows the ground truth. To the best of our knowledge, runtime assumption learning in which information can only be exchanged via peer-to-peer messages has not yet been discussed directly.
Since no component has access to another’s internal state, necessary information must be communicated through explicit messages, which need to contain more than a binary (or ternary) verdict. Instead, a message should explain why a goal cannot be guaranteed, and do it in a way that is sufficient for the receiver to act on. In the current research environment, “explanations” often refer to explainable AI (see surveys such as Arrieta et al. [4] or Longo et al. [18]), and other human-centric concepts like trust and post-hoc blame attribution (see for example Shymanski et al. [25], Wachter et al. [27] or Zehtabi et al. [29]). Our work instead considers machine-to-machine explanations for increased cooperation. A particularly useful family of explanations for this purpose are counterfactual explanations [16], i.e., messages of the form “If ABC was like this, then XYZ could have been achieved.” One may ask whether such messages are explanations or simply constraint propagation. We use the term in the technical sense of Fey et al. [11]: a message is derived from an explanation pattern, it is counterfactual (a minimal change to the belief under which the target holds), and it is actionable for the receiver. In contrast to plain constraint propagation, it carries the reason for the failure, i.e., which assumption is responsible and the maximal range in which it would succeed, rather than just a value or a verdict.
A 2025 survey by Straub et al. [26] reviewed recent papers concerning explanations in self-adaptive systems. Apart from noticing a recent surge of interest in the topic, they identified the lack of actionable explanations as a gap in current research, which is where our paper is located. Other algorithms for self-adaptive systems comprise fault detection and repair, but provide no explanations.
Section 2 introduces background, Section 3 presents the generation of a sufficient, permissive, and actionable counterfactual explanation, Section 4 demonstrates it on a chain, and Section 5 discusses its embedding in a peer-to-peer protocol.
Preliminaries
In our work we use the explanation pattern introduced by Fey et al. [11]. Let be the explanation pattern owned by every component in a world, where is a model of the component and partial model of its surroundings, is a set of constraints defining the (belief about the) current state, and is an explanation target – for example the satisfaction of some specifications. The system tries to derive from and whether is justified, unjustified, or inconclusive.
An assume-guarantee contract (A/G-contract) is a tuple of assertions called the assumption and the guarantee (cf. [5]). The assumption generally refers to conditions of the environment. An implementation satisfies if every environment state that satisfies also satisfies , i.e., .
Let be an unsatisfiable set of clauses. Then a minimal correction subset (MCS) is a set of clauses such that is satisfiable and for every proper subset , is not satisfiable (cf. [23, 17]). An MCS can, for example, be obtained as the complement of a maximum satisfiable subset returned by a MaxSMT solver; such a minimum correction set is in particular minimal.
Given the explanation pattern and the situation that either or , a counterfactual explanation is a minimal difference between and a such that .
Let be a logical formula over the reals and booleans, where and . Quantifier elimination computes a quantifier-free formula , such that holds for every assignment of over . Universal quantifiers are eliminated analogously via .
Explanation Generation
In a system, each component holds assumptions about the environment and guarantees which it must uphold. In terms of the previously introduced explanation pattern, the guarantee corresponds to the explanation target , while the assumption is the set of environment states for which . The explanation generation takes place between two individual components, which we will refer to as the explainer () and the receiver (), whenever notices that its target cannot be fulfilled under the current believed state of its environment. We consider a setting in which operates under partial influence of .
Following the notation of Section 2, maintains a model , which is not subject to repair in this work, and a belief . encodes hard knowledge about ’s own state, while encodes the soft, shared variables whose values are under ’s influence. is therefore a subset of ’s contract assumptions, specifically the subset of changeable parameters.
We distinguish the shared variables constrained by , whose values are controlled by , from all remaining variables . The latter comprise ’s own settings, quantities derived via , and uncontrolled quantities whose current value is known from , e.g., environmental properties.
Given an explanation target such that or , computes and communicates a counterfactual explanation as follows.
Failure detection. evaluates if . If the target is not justified, explanation generation is initiated.
Diagnosis. encodes as a partial MaxSMT problem, treating clauses from , , and as hard, and clauses from as soft. Since only is eligible for correction, the unsatisfied soft clauses returned by the solver form a restricted minimal correction subset . We write for the shared variables occurring in .
Range characterization. The clauses in are removed from , leaving unconstrained by beliefs. then applies quantifier elimination to
yielding a quantifier-free formula over only. Optionally, is simplified to remove redundant clauses and increase interpretability, e.g., using the Z3 tactic propagate-ineqs.
Communication. The counterfactual explanation is sent to .
Response. adopts as a hard constraint in its own belief set , attempts to satisfy its own target under this new restriction, and returns either a rejection or a new value for satisfying .
Since all uncontrolled quantities have known values, the only freedom left in lies in ’s own settings. A satisfying assignment therefore describes a configuration that can actually realise, which is the guarantee needs from . Equation 1 encodes this relationship for the explanation formula which must be equivalent to the existence of an assignment of satisfying the conjunction of model, hard belief, target, and the soft belief without . Quantifier elimination preserves this relationship. If uncontrolled quantities were only known up to bounds, the target would have to hold for all admissible values, requiring a universal quantifier in and a diagnosis step consistent with it. We leave this extension to future work. We chose to use SMT for its expressiveness and mature tool support. Other formalisms such as game graphs (cf. [7]) have concepts similar to MCSs and should allow an adaptation.
Example Consider the example shown in Figure 1. Wind turbine is located downstream of wind turbine , which causes it to miss its energy production target because of ’s wake effect. ’s model consists of the physical equations

Figure 1. Wind turbines: is located in the wake of .
where is the wind speed and is ’s fixed rotor efficiency. ’s target is . The wake factor is encoded as a soft constraint in , while contains the wind speed and . Hence , while consists of known values and derived quantities.
Failure detection. evaluates its model under :
Since , and initiates explanation generation.
Diagnosis. encodes as a partial MaxSMT problem with hard part and soft part :
The solver returns the restricted MCS .
Range characterization. Removing from leaves free. Eliminating and (and the known , ) yields , which simplifies to :
Communication in a Chain
In this section, we present an executable proof-of-concept of the explanation exchange on a chain topology, where each component is influenced by exactly one other component and influences exactly one component, excluding the two end components. In a chain, Step 5 recurses: a receiver that cannot satisfy its own target under the received constraint starts Step 1 itself.1
Figure 2 shows three components in a chain topology. Component A is influenced only by component B, which is influenced only by component C. The turbine’s (A) target is to keep the energy within 1 unit of the production goal , which is inherited from the controller (B). B’s target is to balance the energy production goal and the projected energy consumption , which it inherits from a model household (C). C tries to maximize the number of active appliances, under the constraint that it needs to activate at least 2. The appliances 1, 2, and 3 consume 2, 3, and 1.5 energy units each. Non-negativity clauses are omitted; are booleans, all other variables are reals.
!Simplified energy grid example showing Turbine (A), Controller (B), and Household (C). Turbine (A): Target: ; Model: ; Influence: . Controller (B): Target: ; Model: ; Influence: . Household (C): Target: ; Model: bbox=0.13,0.09,0.87,0.25
Figure 2. Simplified energy grid example. denotes the current wind speed. “Influence: ” means that is controlled by and currently has the value 6.5. The booleans count as 0/1 in sums.
At wind speed 7, this system functions without problem, since A can meet the production target for all 3 appliances. However, when the wind speed drops, then first the production goal and finally the appliance schedule need to be adjusted.
Figure 3 shows abridged logs. At wind speed 6, A sends a message to B, requesting a reduction of the production goal. B can accommodate the request without involving C and sends back . However, at wind speed 3, B cannot accommodate the request on its own, so it sends a request to C, asking it to reduce its projected consumption. C can accommodate the request by turning off one of its appliances, and sends a new value back, , which B uses to lower the production goal to . At wind speed 1, every pair of appliances exceeds the requested consumption, so the exchange results in a failure.
Figure 3. Abridged logs of the chain explanation exchange
Since each component only has at most one upstream and one downstream neighbor, our explanation generation is sufficient to serve as a minimalistic exchange protocol. As presented by Luckcuck et al. [19], in that case each assume and guarantee are connected by the additional rule , which we integrate by adding the explanation to the receiver’s model check. Since a chain has neither cycles nor branches, messages cannot propagate endlessly, therefore any exchange ends after at most messages, where is the number of components. If a component possesses multiple shared variables, there may also be multiple candidates for in the diagnosis step, and as a result multiple attempts to solve the same conflict if one exchange fails, but since there are finitely many shared variables, there can only be finitely many subsets . Termination is guaranteed with either a positive result if a satisfying assignment for every component is found, or negative if no attempt results in one.
Discussion
Section 4 demonstrated that our explanation generation approach is viable for one-dimensional chains of components, but additional mechanisms are needed to guarantee termination in more complex networks. In this section, we discuss concrete challenges and possible future directions for the two main features networks can have: branches and cycles.
Branching Topologies (tree, star) In reality, components generally have more than a single connection. This has two effects: First, each additional parent increases the number of possible correction sets, which have to be searched for a working solution by a depth-first search. Second, each child may be affected by a change and start its own exchange. To compensate, our protocol must contain mechanisms to search through different candidates to find a solution, as well as consider side effects before committing.
Cyclic Topologies (ring) In cyclic topologies, a clear hierarchy cannot be established, which means we cannot rely on a root node which makes the last decision to guarantee termination. Without precautions, a message may infinitely cycle through the network. There are two approaches we are considering at this moment: Firstly, if the protocol could guarantee iterative tightening of constraints, then the exchange will eventually end with either a successful feedback, or a failure when a component determines that tightened constraints exclude any solution (cf. Abstraction Refinement [14]). Secondly, if the explanation carries additional information about each involved component, then perhaps we can avoid having to enter a cycle. Once a component would attempt to contact a previously involved component, it could instead directly determine whether the solution is viable from the included constraints.
Conclusion In this research preview paper, we presented a generation method for counterfactual explanations from formal specifications using partial MaxSMT and quantifier elimination. We demonstrated how an exchange of these explanations can allow a simple system of components to organize without the need for a central decision maker. In the future, we would like to embed the explanation generation in a protocol which works in a wider range of topologies as illustrated above. Luckcuck et al. [19] provide rules and inferences for the branching case, which we would like to adapt. For the cyclic case, they discuss the same difficulties and argue that only partial conclusions can be drawn at this point. In this version, we considered precise knowledge, especially about the environment, but in reality that may not be available. Having only bounds on certain parameters would change the formula used for the range characterization, where are variables on which we only have bounds:
Acknowledgement This work has been funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) project no. 513623283 as part of the Research Training Group CAUSE.
AI Usage The code for the demonstration was written with use of Anthropic’s Claude coding assistant.
References
- [1]Martín Abadi & Leslie Lamport (1993): Composing Specifications. ACM Transactions on Programming Languages and Systems 15(1), pp. 73–132, doi:10.1145/151646.151649. Available at https://lamport.azurewebsites.net/pubs/abadi-composing.pdf.DOI
- [2]Ashwani Anand, Kaushik Mallik, Satya Prakash Nayak & Anne-Kathrin Schmuck (2023): Computing Adequately Permissive Assumptions for Synthesis. In: Tools and Algorithms for the Construction and Analysis of Systems (TACAS 2023), LNCS 13994, Springer, pp. 211–228, doi:10.1007/978-3-031-30820-8_15. arXiv:2301.07563.DOI
- [3]Dana Angluin (1987): Learning Regular Sets from Queries and Counterexamples. Information and Computation 75(2), pp. 87–106, doi:10.1016/0890-5401(87)90052-6.
- [4]Alejandro Barredo Arrieta, Natalia Díaz-Rodríguez, Javier Del Ser, Adrien Bennetot, Siham Tabik, Alberto Barbado, Salvador Garcia, Sergio Gil-Lopez, Daniel Molina, Richard Benjamins, Raja Chatila & Francisco Herrera (2020): Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI. Information Fusion 58, pp. 82–115, doi:10.1016/j.inffus.2019.12.012. arXiv:1910.10045.DOI
- [5]Albert Benveniste, Benoit Caillaud, Dejan Nickovic, Roberto Passerone, Jean-Baptiste Raclet, Philipp Reinkemeier, Alberto Sangiovanni-Vincentelli, Werner Damm, Thomas A. Henzinger & Kim G. Larsen (2018): Contracts for System Design. Foundations and Trends in Electronic Design Automation 12(2–3), pp. 124–400, doi:10.1561/1000000053. Available at https://people.rennes.inria.fr/Albert.Benveniste/pub/ContractsMonograph2018.pdf.
- [6]Davide G. Cavezza, Dalal Alrajeh & András György (2020): Minimal Assumptions Refinement for Realizable Specifications. In: Proceedings of the 8th International Conference on Formal Methods in Software Engineering (FormaliSE), pp. 66–76, doi:10.1145/3372020.3391557.DOI
- [7]Krishnendu Chatterjee, Thomas A. Henzinger & Barbara Jobstmann (2008): Environment Assumptions for Synthesis. In: CONCUR 2008 – Concurrency Theory, LNCS 5201, Springer, pp. 147–161, doi:10.1007/978-3-540-85361-9_14. Available at https://arxiv.org/pdf/0805.4167.
- [8]Hanyue Chen, Yu Su, Miaomiao Zhang, Zhiming Liu & Junri Mi (2023): Learning Assumptions for Compositional Verification of Timed Automata. In: Computer Aided Verification (CAV 2023), Part I, LNCS 13964, Springer, pp. 40–61, doi:10.1007/978-3-031-37706-8_3. Available at https://library.oapen.org/handle/20.500.12657/75412.DOI
- [9]Jamieson M. Cobleigh, Dimitra Giannakopoulou & Corina S. Pasareanu (2003): Learning Assumptions for Compositional Verification. In: Tools and Algorithms for the Construction and Analysis of Systems (TACAS 2003), LNCS 2619, Springer, pp. 331–346, doi:10.1007/3-540-36577-X_24. Available at https://ntrs.nasa.gov/api/citations/20030017771/downloads/20030017771.pdf.DOI
- [10]Lu Feng, Marta Kwiatkowska & David Parker (2011): Automated Learning of Probabilistic Assumptions for Compositional Reasoning. In: Fundamental Approaches to Software Engineering (FASE 2011), LNCS 6603, Springer, pp. 2–17, doi:10.1007/978-3-642-19811-3_2. Available at https://www.prismmodelchecker.org/papers/fase11.pdf.DOI
- [11]Goerschwin Fey, Martin Fränzle & Rolf Drechsler (2022): Self-Explanation in Systems of Systems. In: 2022 IEEE 30th International Requirements Engineering Conference Workshops (REW), IEEE, pp. 85–91, doi:10.1109/REW56159.2022.00023.DOI
- [12]Ferdinando Fioretto, Enrico Pontelli & William Yeoh (2018): Distributed Constraint Optimization Problems and Applications: A Survey. Journal of Artificial Intelligence Research 61, pp. 623–698, doi:10.1613/jair.5565. arXiv:1602.06347.DOI
- [13]Hadar Frenkel, Orna Grumberg, Corina S. Pasareanu & Sarai Sheinvald (2022): Assume, Guarantee or Repair: A Regular Framework for Non-Regular Properties. International Journal on Software Tools for Technology Transfer 24(5), pp. 667–689, doi:10.1007/s10009-022-00669-9.DOI
- [14]Mihaela Gheorghiu Bobaru, Corina S. Pasareanu & Dimitra Giannakopoulou (2008): Automated Assume-Guarantee Reasoning by Abstraction Refinement. In: Computer Aided Verification (CAV 2008), LNCS 5123, Springer, pp. 135–148, doi:10.1007/978-3-540-70545-1_14.
- [15]Dimitra Giannakopoulou, Corina S Păsăreanu & Howard Barringer (2005): Component verification with automatically generated assumptions. Automated Software Engineering 12(3), pp. 297–320, doi:10.1007/s10515-005-2641-y.DOI
- [16]David K. Lewis (1973): Counterfactuals. Blackwell / Harvard University Press. Revised edition, Wiley-Blackwell, 2001, ISBN 978-0-631-22495-2.
- [17]Mark H. Lifton & Karem A. Sakallah (2008): Algorithms for Computing Minimal Unsatisfiable Subsets of Constraints. Journal of Automated Reasoning 40(1), pp. 1–33, doi:10.1007/s10817-007-9084-z.
- [18]Luca Longo, Mario Bracc, Federico Cabitza, Jaesik Choi, Roberto Confalonieri, Javier Del Ser, Riccardo Guidotti, Yoichi Hayashi, Francisco Herrera, Andreas Holzinger, Richard Jiang, Hassan Khosravi, Freddy Lecue, Giancarlo Malgieri, Andrés Páez, Wojciech Samek, Johannes Schneider, Timo Speith & Simone Stumpf (2024): Explainable Artificial Intelligence (XAI) 2.0: A Manifesto of Open Challenges and Interdisciplinary Research Directions. Information Fusion 106:102301, doi:10.1016/j.inffus.2024.102301.DOI
- [19]Matt Luckcuck, Marie Farrell, Angelo Ferrando, Rafael C. Cardoso, Louise A. Dennis & Michael Fisher (2026): A Compositional Approach to Verifying Modular Robotic Systems. Robotics and Autonomous Systems 205, p. 105648, doi:10.1016/j.robot.2026.105648. arXiv:2208.05507v2.DOI
- [20]Jayadev Misra & K. Mani Chandy (1981): Proofs of Networks of Processes. IEEE Transactions on Software Engineering 7(4), pp. 417–426, doi:10.1109/TSE.1981.230844.DOI
- [21]Chunyan Mu & Muhammad Najib (2026): Counterfactual Reasoning for Causal Responsibility Attribution in Probabilistic Multi-Agent Systems. arXiv preprint. arXiv:2605.13077.arxiv.org/abs/2605.13077
- [22]Timothy Parker, Umberto Grandi & Emiliano Lorini (2023): Anticipating Responsibility in Multiagent Planning. In: Proceedings of the 26th European Conference on Artificial Intelligence (ECAI 2023), Frontiers in Artificial Intelligence and Applications 372, pp. 1859–1866, doi:10.3233/FAIA230474. arXiv:2307.16685.DOI
- [23]Raymond Reiter (1987): A Theory of Diagnosis from First Principles. Artificial Intelligence 32(1), pp. 57–95, doi:10.1016/0004-3702(87)90062-2.DOI
- [24]Alberto Sangiovanni-Vincentelli, Werner Damm & Roberto Passerone (2012): Taming Dr. Frankenstein: Contract-Based Design for Cyber-Physical Systems. European Journal of Control 18(3), pp. 217–238, doi:10.3166/ejc.18.217-238. Available at https://disi.unitn.it/~roby/pdfs/ASVDammPasserone12EJC.pdf.
- [25]Joe Shymanski, Jacob Brue & Sandip Sen (2026): Beyond Satisfaction: From Placebic to Actionable Explanations for Enhanced Understandability. In Davide Calvaresi et al., editors: Explainable, Trustworthy, and Responsible AI and Multi-Agent Systems (EXTRAAMAS 2025), LNCS 15936, Springer, doi:10.1007/978-3-032-01399-6_10. arXiv:2512.06591.DOI
- [26]Raphael Straub, Florian Sihler, Ali Torbati, Cong Wang, Raffaela Groner, Verena Klös & Matthias Tichy (2026): Explainability in Self-Adaptive Systems: A Systematic Literature Review. In: 51st Euromicro Conference on Software Engineering and Advanced Applications (SEAA 2025), Part II, LNCS 16082, Springer, pp. 280–297, doi:10.1007/978-3-032-04200-2_19.DOI
- [27]Sandra Wachter, Brent Mittelstadt & Chris Russell (2018): Counterfactual Explanations without Opening the Black Box: Automated Decisions and the GDPR. Harvard Journal of Law & Technology 31(2), pp. 841–887, doi:10.2139/ssrn.3063289. arXiv:1711.00399.DOI
- [28]Makoto Yokoo, Edmund H. Durfee, Toru Ishida & Kazuhiro Kuwabara (1998): The Distributed Constraint Satisfaction Problem: Formalization and Algorithms. IEEE Transactions on Knowledge and Data Engineering 10(5), pp. 673–685, doi:10.1109/69.729707.DOI
- [29]Parisa Zehtabi, Alberto Pozanco, Ayala Bloch, Daniel Borrajo & Sarit Kraus (2024): Contrastive Explanations of Centralized Multi-Agent Optimization Solutions. Proceedings of the International Conference on Automated Planning and Scheduling 34(1), pp. 671–679, doi:10.1609/icaps.v34i1.31530. arXiv:2308.05984.DOI