CONSENSUS Project: Identifying publicly acceptable policy implementations

Even though it is unrealistic to expect citizens to pinpoint the policy implementation that they prefer from the set of alternatives, it is still possible to infer such information through an exercise of ranking the importance of policy objectives according to their opinion. Assuming that the mapping between policy options and objective evaluations is a priori known (through models and simulations), this can be achieved either implicitly through appropriate analysis of social media content related to the policy objective in question or explicitly through the direct feedback provided in the frame of a game. This document focuses on the presentation of a policy model, which reduces the policy to a multi-objective optimization problem and mitigates the shortcoming of the lack of social objective functions (public opinion models) with a black-box, games-for-crowds approach.

💡 Research Summary

The paper introduces a novel framework for incorporating public preferences into policy design by treating policy formulation as a multi‑objective optimization problem and extracting citizens’ implicit weightings of policy objectives. Recognizing that it is unrealistic to expect individuals to directly select from a large set of concrete policy alternatives, the authors propose to infer preferences from how citizens rank the importance of underlying objectives. A crucial prerequisite is the existence of a pre‑specified mapping between policy options and the quantitative evaluation of each objective, which can be derived from established simulation models (e.g., macro‑economic, energy‑system, or social‑impact models).

Two complementary data‑collection channels are described. The first leverages large‑scale social‑media and news‑comment streams. After standard natural‑language‑processing (tokenization, topic modeling, sentiment analysis), the frequency, polarity, and co‑occurrence of objective‑related terms are fed into a Bayesian network that estimates probabilistic weights for each objective. This “implicit” approach captures broad public discourse but suffers from representativeness bias, language ambiguity, and sentiment‑analysis errors.

The second channel is an explicit “games‑for‑crowds” platform. Participants engage in an interactive game where they are repeatedly presented with pairs or small sets of policy objectives and asked to order them by perceived importance. The game design incorporates multiple rounds, random pairing, and a points‑based reward system to mitigate strategic gaming and selection bias. By carefully balancing incentives and introducing stochastic elements, the authors aim to elicit sincere, high‑resolution weight vectors from a diverse crowd.

Collected weight vectors (w_i) are then used to collapse the multi‑objective function f(x)=∑w_i·g_i(x) into a single scalar objective, or alternatively to guide a Pareto‑front search using evolutionary algorithms such as NSGA‑II. The policy decision space x includes variables such as tax rates, emission caps, and education spending, while each objective g_i(x) represents a societal goal (economic growth, environmental protection, social welfare, etc.). To keep the computational burden tractable, the authors employ meta‑modeling (Gaussian‑process surrogates) that approximate the expensive simulation outputs, allowing rapid evaluation of candidate policies.

A central contribution of the work is the “simulation‑experiment‑feedback” loop. After an initial set of weights is inferred, the optimization yields a set of candidate policies. These are tested in pilot simulations or limited real‑world trials, and the observed outcomes are fed back to refine both the policy‑objective mapping and the inferred public weights. This iterative process helps to correct for model misspecification and ensures that the final policy recommendations remain aligned with evolving public sentiment.

The paper also discusses several technical challenges. Non‑linear interactions among objectives can cause the mapping to be highly sensitive, demanding robust surrogate models. Social‑media data may be dominated by vocal minorities, requiring stratified sampling and weighting schemes to achieve demographic balance. In the game setting, participants might adopt strategic behavior to maximize scores rather than reveal true preferences; the authors mitigate this through randomization and by limiting the visibility of others’ rankings. Moreover, the computational cost of generating a dense Pareto front grows exponentially with the number of objectives, prompting the use of dimensionality‑reduction techniques and adaptive sampling.

Limitations acknowledged include the static treatment of objective interdependencies (the current Bayesian network does not capture temporal dynamics) and the reliance on pre‑existing simulation models, which may embed expert biases. Future research directions propose extending the framework with dynamic Bayesian networks, reinforcement‑learning‑based adaptive weight estimation, and real‑time feedback mechanisms within the game to continuously update preferences as participants interact.

In summary, the CONSENSUS project offers a comprehensive, data‑driven methodology that bridges the gap between expert‑centric policy modeling and citizen‑centric preference elicitation. By converting policy design into a multi‑objective optimization problem and employing both implicit (social‑media) and explicit (crowd‑gaming) channels to infer objective weights, the approach promises more democratically legitimate and socially acceptable policy outcomes, while also providing a systematic loop for continual model improvement.