When are decisions made? The answer to this question may be obvious. Usually, decisions are made just in time. For example, we choose a restaurant just before going out for lunch. Or we choose a product just when we want to buy it. Making a decision earlier does not make much sense as we don’t know all the available alternatives in advance.

Perhaps we should reframe the question as follows: when are we crafting the method for making the decision, i.e. the specific method that will narrow down the alternatives to a single one? Now the answer is less obvious. Some people may follow habits that they have gathered by education, meaning that the essential decisions governing their lives may have been made even before they have been born. Others may have experimented with different alternatives, compared their outcomes, and derived some rules that they follow. Even others may be more flexible and work with explicit preferences that they are using to compare the available options just in time when they need to make a decision. So there is a large spectrum of possibilities for the moments at which decision are made. The decisions may be made before birth or during life time by crafting a method or just in time by comparing the available alternatives with our preferences.

Interestingly, the same spectrum applies to computational methods for decision making. There are methods that use explicit preference models to determine a best decision in the moment when the decision needs to be made and all available alternatives are known. Next there are reinforcement learning methods that learn a policy by experimentation. And finally, there are systems such as neural networks that apply a predetermined policy learned from data.

This third approach completely separates the moment of crafting the decision-making method and the moment at which a decision is required. Due to this separation in time, those systems cannot take particular characteristics of the situation into account in which a decision is required. This situation may allow for new alternatives that were not known when the method was crafted. Furthermore, some of the anticipated alternatives may not be possible anymore. The method has been crafted with the help of training data that are described by a fixed set of features. There is an implicit assumption that these features uniquely determine the set of alternatives that are available in future.

This argument shows that the standard machine learning approach commits to a deterministic policy too early, namely when it is not known which kind of information may be available at the moment when a decision is required. A more flexible approach is a least commitment method which allows for adjustments later on. So instead of learning a policy that proposes a single choice, a least commitment learner would learn a policy that proposes a set of choices. A concrete choice can then be made just in time when a decision is required. This final choice can be made by some other method that takes additional information into account such as situation-dependent preferences.

This idea of least commitment is quite natural. It was wonderful to learn at the recent DA2PL workshop that it is subject of a new research effort. Before we will give some impressions of DA2PL 2022, let us mention our forthcoming event, namely the 14th Multidisciplinary Workshop on Advances in Preference Handling M-PREF 2023. It will be held in Macao, S.A.R and not in Cape Town, South Africa as previously announced as the hosting conference IJCAI-2023 has changed its location. The workshop will take place on August 21, 2023 and is a physical event, organized by Haris Aziz, Ulrich Junker, Xinhang Lu, Nicholas Mattei, and Andrea Passerini. As in previous years, the workshop will address all computational aspects of preference handling. Submissions are expected by May 1, 2023. Please consult the workshop web site for more information. This workshop will also serve as this year’s meeting of the working group on Advances in Preference Handling.

The DA2PL 2022 workshop has been organized by Sébastien Destercke and Khaled Belahcene in Compiègne, France on November 17 and 18, 2022. It has been the sixth edition of a workshop series that brings together researchers from decision analysis and preference learning. The program consisted of three excellent invited talks as well of thirteen highly interesting regular presentations about topics such as preference learning, ordering of alternatives, space and time in preferences, and explaining preferences.

Thomas Augustin (Ludwig Maximilian University of Munich) gave the first invited talk. He applied insights from decision making under strict uncertainty to machine learning. In decision making, set-based concepts have been developed to precisely model the partial information that is available in presence of uncertainty. These concepts include sets of potential utility functions, sets of prior distributions, and choice sets, i.e. sets of preferred alternatives. This set-based approach gives an answer to the problem of early commitment, which we have explained above. Thomas then applied this approach to a variety of topics in machine learning such as classification. A first idea consists in learning set-based classifiers which predict an optimal set of classes instead of a single class. A second idea consists in learning a whole set of classifiers by using a set of loss functions or a set of prior distributions. As data may be partial, it is also possible to use a set of potential data in order to learn a set of classifiers. In the second part of his talk, Thomas used decision making methods to compare and rank classifiers under multiple criteria (such as accuracy and area under curve) and for different benchmark sets. Thomas used a generalized notion of stochastic dominance for this purpose in order to avoid problems with aggregations.

The second invited talk was given by Nicolas Usunier (Meta AI) about fair allocation in recommender systems. Nicolas explained that recommender systems are learning user preferences in a first step and then allocate ranked lists of items to users while taking the user preferences into account as well as the interests of the providers, who want a good exposure of their items. Nicolas defined an optimization objective for each user and each item and then investigated methods from social choice theory in order to make trade-offs in a way that is fair for both users and providers. Nicolas not only required that allocations are Pareto-efficient, but also Lorenz-efficient. The latter allows for re-allocation of goods from a first user to a second user if the first user has more goods than the second user. Nicolas showed that these two properties can be achieved by applying a strictly monotonic and strictly concave function to each of the objectives and then defining the global welfare as the sum of the results. As global allocation has several computational drawbacks, Nicolas developed an alternative approach based on contextual bandit algorithms that have concave rewards and ensure fairness of item exposure. These algorithms determine a ranked list of items for a user upon demand and integrate preference learning into the ranking task.

Denis Bouyssou (Université Paris Dauphine) gave a tutorial about the conjoint measurement approach to preference modeling for multi-attributed alternatives. Denis recalled the assumptions of the additive value functions. He then discussed measurement in physics where objects can be compared, combined into new objects (e.g. the concatenation of two rods), and standard sequences can be defined. When applying this approach to alternatives that are characterized by multiple attributes, there is no concatenation operation anymore, but new alternatives can be formed by modifying alternatives. The purpose of this operation is to assess tradeoffs between attributes. For example, which gain would be needed on a first attribute in order to compensate for a loss on another attribute? The answer of this question will help to transform an alternative into a new alternative such that the decision maker is indifferent about both alternatives. Denis then carefully explained necessary and sufficient conditions for the existence of an additive value function that represents a weak preference order. This includes preferential independence of the order on all subsets of attributes. After this central result, Denis discussed some misconceptions. In particular, he insisted that a weighted sum must not be confused with the additive model. In the final part of the talk, Denis surveyed some models with interactions between attributes such as polynomial models, generalized additive independence, CP-nets, and Choquet-integrals. A very general model is the decomposable model, which uses an arbitrary aggregation operator that is increasing on all arguments. He finished the talk by discussing models with intransitive preferences as well as sorting problems.

The slides of the invited talks and the papers of the regular presentations can all be found on the workshop website.