Daily Summary for 2026-10-18

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: It's the eighteenth of October, twenty twenty-six, and this is the day's research.

Warren: One new paper came out today.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We'll take the day in one pass, then pull out the paper we're staying with.

Warren: Start us off with Delphos: A reinforcement learning framework for assisting discrete choice model specification.

The day's research: Ruth: Today's work centers on automating and improving discrete choice models.

Warren: That automation reduces the manual effort for modelers in utility functions, correct?

Iris: The main focus is Delphos, a deep reinforcement learning framework.

Ruth: Delphos suggests model specifications through a sequence of decisions.

Warren: What kind of decisions does the framework suggest?

Iris: It suggests specifications based on those sequential choices.

Ruth: Delphos treats building a model as an iterative decision-making process where an agent learns to choose actions like adding specific constants or applying non-linear transformations to attributes.

Warren: Does that mean it's learning those choices through trial and error?

Iris: It uses a Deep Q-Network to figure out which sequence of these specification decisions leads to the best final model.

Ruth: The reward function is set by the human modeller, guiding that learning process.

Lucky paper: 2506.06410: Ruth: So, we're digging into the mechanics of Delphos now, focusing on how this reinforcement learning framework actually functions step-by-step.

Warren: It treats model specification as a Markov Decision Process, which means the current partial specification defines the state and the feasible refinements define the actions.

Iris: That sounds like an MDP where each action is choosing something concrete like adding a new encoded variable or changing an existing one in the model.

Ruth: Exactly, and because it uses a Deep Q-Network to approximate those state-action values, it’s learning which sequence of these modelling actions leads to high quality models.

Warren: What exactly defines that quality signal for the agent?

Iris: The reward function is what translates the modeller's goals into a learning signal, combining things like log-likelihood and McFadden’s pseudo-R2 metrics.

Ruth: So it’s not just about one metric; it combines model fit with behavioral expectations, which is important because that’s where the real utility lies.

Warren: And the paper tests this on both simulated datasets and empirical ones, including benchmarking against metaheuristics like Variable Neighbourhood Search and Simulated Annealing.

Iris: That’s interesting because they show learning curves in simulations where Delphos explores only a small fraction of the feasible modelling space.

Ruth: That suggests it finds effective strategies without needing to try every possible combination of model components, which is a big win for efficiency.

Warren: But what about when the agent gets stuck? The paper mentions that it might initially produce behaviourally implausible models if there are conflicting objectives in the reward function.

Iris: That’s a real caveat; the agent can struggle with those conflicting goals when trying to find a good balance between fit and behavior.

Ruth: The future work section points out offline learning methods where the replay buffer is fed known sequences of actions, which could help stabilize that training process.

Warren: That moves it from just automating the search to learning how to actively assist the specification process itself, which is a significant shift in how we think about these tools.

Iris: It paves the way for a new generation of intelligent tools for discrete choice modelling by moving beyond executing predefined procedures.