Periodic Reporting for period 2 - TUPLES (Trustworthy Planning and Scheduling with Learning and Explanations)
Okres sprawozdawczy: 2024-04-01 do 2025-09-30
Planning and scheduling (P&S) is a core AI area that supports humans in planning, organising, and optimising actions to achieve complex goals. Despite strong demand in industry and public services, current P&S tools lack key trustworthiness properties.
TUPLES was a 3-year Horizon EU project to develop trustworthy P&S systems that are scalable, safe, robust, and explainable.
Our scientific objectives were: (1) integrating symbolic P&S with data-driven methods to combine safety and transparency with modelling power and scalability; (2) developing verification approaches to ensure robustness and safety of sequences of interacting ML-based decisions; and (3) creating explanation methods that justify plans and schedules and support interactive, transparent decision making.
Technologically, our goals were to demonstrate and evaluate our methods in laboratory settings across use cases in manufacturing, flight diversion, sport management, waste collection, and energy management, advancing TRL2 → TRL4, and to release open-source tools and test environments to enable development and assessment of trustworthy P&S systems.
From a business and societal perspective, TUPLES sought to provide strategies and tools to improve Decision Support Systems (DSS) effectiveness in these domains and in similar P&S applications. Beyond efficiency gains, the main legacy is increased trustworthiness: more reliable, scalable, robust, and explainable DSS are expected to accelerate adoption and deliver significant environmental, societal, and economic benefits across the EU.
TUPLES was a 3-year Horizon EU project to develop trustworthy P&S systems that are scalable, safe, robust, and explainable.
Our scientific objectives were: (1) integrating symbolic P&S with data-driven methods to combine safety and transparency with modelling power and scalability; (2) developing verification approaches to ensure robustness and safety of sequences of interacting ML-based decisions; and (3) creating explanation methods that justify plans and schedules and support interactive, transparent decision making.
Technologically, our goals were to demonstrate and evaluate our methods in laboratory settings across use cases in manufacturing, flight diversion, sport management, waste collection, and energy management, advancing TRL2 → TRL4, and to release open-source tools and test environments to enable development and assessment of trustworthy P&S systems.
From a business and societal perspective, TUPLES sought to provide strategies and tools to improve Decision Support Systems (DSS) effectiveness in these domains and in similar P&S applications. Beyond efficiency gains, the main legacy is increased trustworthiness: more reliable, scalable, robust, and explainable DSS are expected to accelerate adoption and deliver significant environmental, societal, and economic benefits across the EU.
We made significant advances on all three scientific objectives.
1. Hybrid (reasoning/learning) methods
Constraint Acquisition, we reduced expert queries by up to 72% using ML guidance (AAAI24), with further reductions via generalisation queries (AAAI25). We also developed preference-learning methods applied in our Manufacturing use case.
Decision-Focused Learning (DFL): We created the first DFL method for planning (ECAI24) and a multi-stage robust optimisation approach that reuses existing deterministic solvers yet matches the solution quality of scenario-based optimisation while being two orders of magnitude faster (KBS24). This was applied to the Energy use case to increase robustness to demand uncertainty.
Learning to Plan: We pioneered graph-learning heuristics (AAAI24, ICAPS24, IJCAI24) that solve 160% more problems than prior learned heuristics and rival or outperform state-of-the-art classical/numeric planners (ICAPS25, NeurIPS24). This was applied to Beluga. We further contributed theoretical and empirical insights into graph learning for planning (AAAI24, ICAPS24, AAAI25).
2. Verification and testing
We developed a neural network (NN) architecture enabling conservative verification with controllable cost and guarantees even out-of-distribution (AAAI25).
We introduced the first correct multiclass verification for decision-tree ensembles (DTE), and a compression method reducing size ensemble size with minimal accuracy loss (AAAI24, ICML25).
We designed two multi-step safety-verification algorithms for NN and DTE policies, achieving 2–5 orders of magnitude speed-ups and showing DTEs can be 3+ orders faster.
We built a policy-safety debugging loop that identifies unsafe runs, fixes unsafe DTE actions, and substantially reduces unsafe behaviour, enabling full verification in multiple domains. It transforms unsafe Beluga policies into provably safe ones.
3. Explainable planning and scheduling
We extended conflict-based explanations to numeric and probabilistic planning and developed faster/approximate methods (ICAPS24, ECAI24, AAAI25), supporting Manufacturing, Beluga, and Flight Diversion.
We built IPEXCO, an explanation-driven interactive planning platform with LLM conversational interface (XAI25), used in Beluga. We devised CPMpy.tools.explain to provide conflict computation, resolution, and stepwise explanations, and used it n Airbus Manufacturing (CP25).
We advanced explanations of ML policies, including distilling GNNs into C2 logic formulas and extending abductive explanations to NN policies (IJCAI23).
We built demonstrators for all use cases, integrating our research with use-case-specific solutions.
* Airbus Manufacturing: workforce allocation and scheduling via CP/PB/MILP and learned-heuristic guidance; includes disruption generation and conflict-based explanation (CP25 best application paper).
* Airbus Beluga Logistics: deterministic and hybrid planning algorithms, incomplete methods matching competition winners, policy verification/testing, and conflict-based explanations (ECAI25).
* Airbus Flight Diversion: hybrid RL/A* route computation, policy testing, and conflict-based explanations for fuel, time, cost, and safety.
* Optit Waste Collection: scalable 2-stage city-scale CVRP solution with interactive “destroy and repair” exploration, contrastive explanations, and robustness simulation.
* Optit Energy Management: robust Unit Commitment planner using DFL, without needing access to the optimization problem or the solver state; applied to both Optit and simplified planners.
* SciSports Squad Management: robust squad-planning tool integrating verification and tree-compression techniques, deployed in an end-user environment for safe, transparent use.
We released the TUPLES toolkit, aimed at assisting the design and development of trustworthy DSS. It includes:
* Self-Assessment Tool for trustworthiness evaluation;
* TUPLES Lab Python package with robust simulation environments;
* GitLab repository with code for code platforms and demonstrators;
* Contributions to Scikit-Decide, enhancing planning, scheduling, and RL.
We organised a scalability/explainability competition on Beluga, providing generators, simulators, evaluators, baseline solvers, and a new competition platform.
Most technologies advanced to TRL4, some demonstrators TRL5–7, and the competition platform TRL7–8.
1. Hybrid (reasoning/learning) methods
Constraint Acquisition, we reduced expert queries by up to 72% using ML guidance (AAAI24), with further reductions via generalisation queries (AAAI25). We also developed preference-learning methods applied in our Manufacturing use case.
Decision-Focused Learning (DFL): We created the first DFL method for planning (ECAI24) and a multi-stage robust optimisation approach that reuses existing deterministic solvers yet matches the solution quality of scenario-based optimisation while being two orders of magnitude faster (KBS24). This was applied to the Energy use case to increase robustness to demand uncertainty.
Learning to Plan: We pioneered graph-learning heuristics (AAAI24, ICAPS24, IJCAI24) that solve 160% more problems than prior learned heuristics and rival or outperform state-of-the-art classical/numeric planners (ICAPS25, NeurIPS24). This was applied to Beluga. We further contributed theoretical and empirical insights into graph learning for planning (AAAI24, ICAPS24, AAAI25).
2. Verification and testing
We developed a neural network (NN) architecture enabling conservative verification with controllable cost and guarantees even out-of-distribution (AAAI25).
We introduced the first correct multiclass verification for decision-tree ensembles (DTE), and a compression method reducing size ensemble size with minimal accuracy loss (AAAI24, ICML25).
We designed two multi-step safety-verification algorithms for NN and DTE policies, achieving 2–5 orders of magnitude speed-ups and showing DTEs can be 3+ orders faster.
We built a policy-safety debugging loop that identifies unsafe runs, fixes unsafe DTE actions, and substantially reduces unsafe behaviour, enabling full verification in multiple domains. It transforms unsafe Beluga policies into provably safe ones.
3. Explainable planning and scheduling
We extended conflict-based explanations to numeric and probabilistic planning and developed faster/approximate methods (ICAPS24, ECAI24, AAAI25), supporting Manufacturing, Beluga, and Flight Diversion.
We built IPEXCO, an explanation-driven interactive planning platform with LLM conversational interface (XAI25), used in Beluga. We devised CPMpy.tools.explain to provide conflict computation, resolution, and stepwise explanations, and used it n Airbus Manufacturing (CP25).
We advanced explanations of ML policies, including distilling GNNs into C2 logic formulas and extending abductive explanations to NN policies (IJCAI23).
We built demonstrators for all use cases, integrating our research with use-case-specific solutions.
* Airbus Manufacturing: workforce allocation and scheduling via CP/PB/MILP and learned-heuristic guidance; includes disruption generation and conflict-based explanation (CP25 best application paper).
* Airbus Beluga Logistics: deterministic and hybrid planning algorithms, incomplete methods matching competition winners, policy verification/testing, and conflict-based explanations (ECAI25).
* Airbus Flight Diversion: hybrid RL/A* route computation, policy testing, and conflict-based explanations for fuel, time, cost, and safety.
* Optit Waste Collection: scalable 2-stage city-scale CVRP solution with interactive “destroy and repair” exploration, contrastive explanations, and robustness simulation.
* Optit Energy Management: robust Unit Commitment planner using DFL, without needing access to the optimization problem or the solver state; applied to both Optit and simplified planners.
* SciSports Squad Management: robust squad-planning tool integrating verification and tree-compression techniques, deployed in an end-user environment for safe, transparent use.
We released the TUPLES toolkit, aimed at assisting the design and development of trustworthy DSS. It includes:
* Self-Assessment Tool for trustworthiness evaluation;
* TUPLES Lab Python package with robust simulation environments;
* GitLab repository with code for code platforms and demonstrators;
* Contributions to Scikit-Decide, enhancing planning, scheduling, and RL.
We organised a scalability/explainability competition on Beluga, providing generators, simulators, evaluators, baseline solvers, and a new competition platform.
Most technologies advanced to TRL4, some demonstrators TRL5–7, and the competition platform TRL7–8.
Technology: Despite targeting TRL4, multiple results are influencing partners’ investment plans; some are already being industrialised. Optit considers Energy and Waste results for upcoming investments; Airbus is industrialising conflict-based explanations and considers TUPLES explainability tools essential for future AI DSS; SciSports continues developing squad-management and other trustworthy-AI applications. Sustained industrial interest fosters post-project collaborations and broader adoption.
Competitiveness: Improved DSS enhance partners’ competitiveness via cost reduction and performance gains. Optit’s Competition Platform is emerging as a market offering, with potential new revenue, dedicated staff, and scale-up opportunities.
Economic: Even small gains in DSS trustworthiness can significantly increase adoption, producing measurable impacts in millions of euros.
Societal: Trustworthy P&S improves efficiency, reducing fuel use, pollution, and CO2 emissions. Enhanced robustness, scalability, safety, and explainability improve user experience, reduce stress, and support productivity and well-being.
Dissemination: 68 publications (75% top-tier), 21 more in preparation; two best-paper recognition. Organised 19 workshops, 5 tutorials, a summer school, and gave 30 invited talks. Beluga competition attracted 19 participants; Linköping University developed an advanced course based on its resources.
Competitiveness: Improved DSS enhance partners’ competitiveness via cost reduction and performance gains. Optit’s Competition Platform is emerging as a market offering, with potential new revenue, dedicated staff, and scale-up opportunities.
Economic: Even small gains in DSS trustworthiness can significantly increase adoption, producing measurable impacts in millions of euros.
Societal: Trustworthy P&S improves efficiency, reducing fuel use, pollution, and CO2 emissions. Enhanced robustness, scalability, safety, and explainability improve user experience, reduce stress, and support productivity and well-being.
Dissemination: 68 publications (75% top-tier), 21 more in preparation; two best-paper recognition. Organised 19 workshops, 5 tutorials, a summer school, and gave 30 invited talks. Beluga competition attracted 19 participants; Linköping University developed an advanced course based on its resources.