Periodic Reporting for period 2 - CHAT-Opt (Conversational Human-Aware Technology for Optimisation)
Reporting period: 2022-12-01 to 2024-05-31
Industry and society are increasingly automating processes, which requires solving constrained optimisation problems. This includes vehicle routing problems in transportation, scheduling of tasks in electricity demand-response programs, rostering of personnel and more. However, the solutions found by state-of-the-art constraint solvers are often found to not match the expectations of domain experts, which reduces the potential of the technology.
This project is based on the idea that we must shift away from the concept of the optimal solution, and must move towards the desired solution. To make the solutions more humane, we must make the solution process more human-aware. The key to this is to integrate techniques from machine learning that can learn about the context of the environment and user; and to make the solution process more explainable and bi-directional.
This project aims to fundamentally advance the state-of-the-art on three fronts, with the following objectives:
- to learn from the environment and the context in which the optimisation operates, by integrating the learning and the predictive models into the optimisation;
- to learn the implicit preferences from the user by letting her/him interact with the solutions thereby learning constraints and preferences; and ultimately
- to develop a conversational constraint solver that can answer unambiguous questions of the users in terms of explanations and alternatives.
The end goal is a system that lets the user co-create a solution that she/he desires, and that this system can adapt to changing needs and preferences over time. The keystone technology is our newly developed CPMpy constraint programming and modeling library.
This project is based on the idea that we must shift away from the concept of the optimal solution, and must move towards the desired solution. To make the solutions more humane, we must make the solution process more human-aware. The key to this is to integrate techniques from machine learning that can learn about the context of the environment and user; and to make the solution process more explainable and bi-directional.
This project aims to fundamentally advance the state-of-the-art on three fronts, with the following objectives:
- to learn from the environment and the context in which the optimisation operates, by integrating the learning and the predictive models into the optimisation;
- to learn the implicit preferences from the user by letting her/him interact with the solutions thereby learning constraints and preferences; and ultimately
- to develop a conversational constraint solver that can answer unambiguous questions of the users in terms of explanations and alternatives.
The end goal is a system that lets the user co-create a solution that she/he desires, and that this system can adapt to changing needs and preferences over time. The keystone technology is our newly developed CPMpy constraint programming and modeling library.
We developed the new and generic CPMpy library, a Constraint Programming and Modeling library in Python. Its novelties include being based on numpy for easy machine learning integration, and allowing for incremental and direct solver access. CPMpy includes many examples, documentation and tests. It can translate to CP, ILP, SMT and SAT solvers in an incremental way, and has boosted our research on each of the objectives.
For the objective of learning from the environment, we have contributed a number of novel techniques for “Decision-Focused Learning”, which is a methodology in which machine learning models are trained based on the effect that the predictions have on the solutions of the solver. We have also created the most comprehensive survey on these techniques to date, titled “Decision-focused learning: Foundations, state of the art, benchmark and future opportunities”, as well as hosting a summer school on the integration of machine learning and constraint solving.
For learning implicit preferences we have made significant advances in passively learning constraints from data for generalized constraint models and MaxSAT models. We also enabled for the first time the use of generic constraint solvers in active constraint acquisition techniques, and were able to reduce the number of interactions by up to 60%. Furthermore we published a methodology for learning implicit preferences for the use case of vehicle routing, the ‘learn-and-route’ approach, which is compatible with all existing vehicle routing software.
As to conversational solving, our initial focus has been on explainable constraint solving. We have published a number of new techniques for explaining why there is no solution or why certain decisions must be taken, as well as techniques to simplify such explanations. We presented an overview of a wide range of explanation techniques, implemented in CPMpy, at a well-appreciated tutorial on explainable constraint programming at the CP23 conference.
An early demonstration of what is possible when all of this is combined is available in our Sudoku Assistant Android app, which allows taking a picture of a paper Sudoku; uses machine learning to reason over what information is present in the image and can show explanations for what the next easiest step is, in case a user is stuck. This work received the AAAI 2023 Best Demonstration Award during the Thirty-Seventh AAAI Conference on Artificial Intelligence in Washington DC, USA.
For the objective of learning from the environment, we have contributed a number of novel techniques for “Decision-Focused Learning”, which is a methodology in which machine learning models are trained based on the effect that the predictions have on the solutions of the solver. We have also created the most comprehensive survey on these techniques to date, titled “Decision-focused learning: Foundations, state of the art, benchmark and future opportunities”, as well as hosting a summer school on the integration of machine learning and constraint solving.
For learning implicit preferences we have made significant advances in passively learning constraints from data for generalized constraint models and MaxSAT models. We also enabled for the first time the use of generic constraint solvers in active constraint acquisition techniques, and were able to reduce the number of interactions by up to 60%. Furthermore we published a methodology for learning implicit preferences for the use case of vehicle routing, the ‘learn-and-route’ approach, which is compatible with all existing vehicle routing software.
As to conversational solving, our initial focus has been on explainable constraint solving. We have published a number of new techniques for explaining why there is no solution or why certain decisions must be taken, as well as techniques to simplify such explanations. We presented an overview of a wide range of explanation techniques, implemented in CPMpy, at a well-appreciated tutorial on explainable constraint programming at the CP23 conference.
An early demonstration of what is possible when all of this is combined is available in our Sudoku Assistant Android app, which allows taking a picture of a paper Sudoku; uses machine learning to reason over what information is present in the image and can show explanations for what the next easiest step is, in case a user is stuck. This work received the AAAI 2023 Best Demonstration Award during the Thirty-Seventh AAAI Conference on Artificial Intelligence in Washington DC, USA.
Research on all three objectives will continue, with as end-goal an intelligent, conversational constraint solver that can provide explanations and suggest smart alternatives in response to user interaction.
With respect to learning from the environment we are focusing on richer architectures, scaling up learning as well as techniques that can learn both parameters in the objective function and the constraints. For example, we are developing use cases in vehicle routing, as well as adapting the techniques to problems in stochastic optimisation.
For the learning of implicit preferences, we are continuing our development of constraint acquisition techniques. More specifically we aim to further develop probabilistic learning techniques that can guide the query generation process, e.g. use machine learning to estimate the probability of constraints being present, combined with constraint acquisition techniques to confirm the presence of hard constraints. For learning the objective function through user interaction, we are looking at a further adaptation of techniques from structured output prediction to the setting of combinatorial optimisation, with the goal of developing a generic methodology for learning the objective function through user interaction.
We will also further develop techniques for explainable constraint solving, with a focus on expressivity on the one hand and scalability on the other hand.
Thanks to recent breakthroughs in the development of instruction-following large language models, we are beginning to explore the possibility of natural language-based interaction in constraint acquisition. This could enable an integrated conversational based constraint solving approach that surpasses the initial goals of the project.
With respect to learning from the environment we are focusing on richer architectures, scaling up learning as well as techniques that can learn both parameters in the objective function and the constraints. For example, we are developing use cases in vehicle routing, as well as adapting the techniques to problems in stochastic optimisation.
For the learning of implicit preferences, we are continuing our development of constraint acquisition techniques. More specifically we aim to further develop probabilistic learning techniques that can guide the query generation process, e.g. use machine learning to estimate the probability of constraints being present, combined with constraint acquisition techniques to confirm the presence of hard constraints. For learning the objective function through user interaction, we are looking at a further adaptation of techniques from structured output prediction to the setting of combinatorial optimisation, with the goal of developing a generic methodology for learning the objective function through user interaction.
We will also further develop techniques for explainable constraint solving, with a focus on expressivity on the one hand and scalability on the other hand.
Thanks to recent breakthroughs in the development of instruction-following large language models, we are beginning to explore the possibility of natural language-based interaction in constraint acquisition. This could enable an integrated conversational based constraint solving approach that surpasses the initial goals of the project.