Generative AI-assisted Participatory Modeling in Socio-Environmental Planning under Deep Uncertainty

Zhihao Pei; Nir Lipovetzky; Angela M. Rojas-Arevalo; Fjalar J. de Haan; Enayat A. Moallemi

cs.AIMar 17, 2026

Generative AI-assisted Participatory Modeling in Socio-Environmental Planning under Deep Uncertainty

Authors: Zhihao Pei, Nir Lipovetzky, Angela M. Rojas-Arevalo, Fjalar J. de Haan, Enayat A. Moallemi

Organizations: School of Computing and Information Systems, Faculty of Engineering and Information Technology, The University of Melbourne, Parkville VIC 3010, Australia · The Commonwealth Scientific and Industrial Research Organization (CSIRO), Research Way, Clayton VIC 3168, Australia · Department of Psychiatry, The University of Melbourne, Parkville VIC 3010, Australia

Abstract

Socio-environmental planning under deep uncertainty requires researchers to identify and conceptualize problems before exploring policies and deploying plans. In practice and model-based planning approaches, this problem conceptualization process often relies on participatory modeling to translate stakeholders' natural-language descriptions into a quantitative model, making this process complex and time-consuming. To facilitate this process, we propose a templated workflow that uses large language models for an initial conceptualization process. During the workflow, researchers can use large language models to identify the essential model components from stakeholders' intuitive problem descriptions, explore their diverse perspectives approaching the problem, assemble these components into a unified model, and eventually implement the model in Python through iterative communication. These results will facilitate the subsequent socio-environmental planning under deep uncertainty steps. Using ChatGPT 5.2 Instant, we demonstrated this workflow on the lake problem and an electricity market problem, both of which demonstrate socio-environmental planning problems. In both cases, acceptable outputs were obtained after a few iterations with human verification and refinement. These experiments indicated that large language models can serve as an effective tool for facilitating participatory modeling in the problem conceptualization process in socio-environmental planning.

Figures & tables

Panel of Figure 1 — Figure 1: The two DMDU processes covered in this paper.

Figure 2: The LLM-assisted workflow for communication with LLMs for initial problem conceptualization. In this diagram, nodes represent discrete steps in the workflow. Red text denotes the information transferred between steps, whereas blue text denotes iterative human validation, verification and refinement of the responses.

Figure 3: The overall experimental design. In our experiments, we considered two case studies, the lake problem and the electricity market problem. For the former, we only considered the sub-case with a comprehensive problem description. For the latter, we considered both the sub-case with a comprehensive problem description and the sub-case with a brief problem description. For each sub-case, we ran the initial problem conceptualization process several times following the proposed workflow.

Types	Parameters	Descriptions
State variables	$X_{t}$	Lake pollution at time $t$ .
Decision variables	$a_{t}$	Rate of anthropogenic pollution at time $t$ .
Transition functions	$f(X_{t})$	System dynamics from $X_{t}$ to $X_{t+1}$ .
Stochastic variables	$\epsilon_{t}$	Rate of natural pollution at time $t$ .
Constants	$\alpha$	Benefit-to-pollution ratio.
Deep uncertainties	$(\mu,\sigma)$ ; $b$ ; $q$ ; $\delta$	Mean and standard deviation of the log-normal distribution for natural inflows; Natural removal rate; Recycling exponent; Discount factor.

Table 1: The key components specified in the baseline description of the lake problem.

Types	Parameters	Descriptions
Decision variables	$(b_{wt},p_{wt})$	Bid for the wind-power producer at time $t$ .
Transition functions	$h_{t}(\cdot)$	Specification on how the market operator determines the market-clearing price and dispatches the schedule.
Stochastic variables	$D_{t}$ ; $p_{it}$ ; $p_{st}$ ; $G_{t}$	Market demand at time $t$ ; Bid price for the conventional producer $i$ at time $t$ ; Bid price for the solar-power producer $s$ at time $t$ ; Actual wind-power production at time $t$ .
Constants	$(\mu_{D},\sigma_{D})$ ; $b_{i}$ ; $(a,b)$ ; $(\mu_{G},\sigma_{G})$ ; $q_{u}$	Mean and standard deviation of market demands; Bid quantity for the conventional producer $i$ ; Solar-power production parameters; Mean and standard deviation of wind-power production; Shortfall penalty coefficient.
Deep uncertainties	$(\mu_{pi},\sigma_{pi})$ ; $(\mu_{ps},\sigma_{ps})$	Mean and standard deviation of bid prices for the conventional producer $i$ ; Mean and standard deviation of bid prices for the solar-power producer $s$ .
Objective functions	$\Pi$	Daily expected revenue.

Table 2: The key components specified in the baseline description of the electricity market problem.

Variables	Descriptions	Formulas
Lake pollution	Lake pollution at time $t$ .	$X_{t}$

Table 3: Time series of variables used to evaluate the consistency in the performance of final Python implementations in the lake problem.

Variables	Descriptions	Formulas
Clearing price	Clearing price at time $t$ .	$c_{t}$
Actual dispatched wind energy	Amount of wind energy actually dispatched at time $t$ .	$x_{wt}=b_{wt}$ if $p_{wt}\leq c_{t}$ otherwise $0$
Wind-power revenue	Revenue of the wind-power producer at time $t$ . Here, $b_{wt}$ and $p_{wt}$ are the bid quantity and price of the wind-power producer; $u_{t}=max(0,x_{wt}-G_{t})$ .	$f(revenue)=c_{t}x_{wt}-q_{u}u_{t}$

Table 4: Time series of variables used to evaluate the consistency in the performance of final Python implementations in the electricity market problem.

Parameters	Descriptions	Values (dimensionless)
$b$	Pollution removal rate	0.42
$q$	Recycling exponent	2.0
$\mu$	Mean of natural pollution inflows	0.02
$\sigma$	Standard deviation of natural pollution inflows	0.0017
$\delta$	Discount factor	0.98
$a_{t}$	Anthropogenic pollution emission at time $t$	Fixed sequence uniformly sampled from [0.02, 0.04]

Table 5: Experimental settings used to generate the time series in the lake problem.

Parameters	Descriptions	Values
$\mu_{D}$	Mean of market demands	800 (MWh)
$\sigma_{D}$	Standard deviation of market demands	200 (MWh)
$b_{i}$	Bid quantity for the conventional producer $i$	[300, 250, 1000] (MWh)
$\mu_{pi}$	Mean of bid prices for the conventional producer $i$	[45, 50, 60] ($/MWh)
$\sigma_{pi}$	Standard deviation of bid prices for the conventional producer $i$	[5, 5, 5] ($/MWh)
$(a,b)$	Solar-power production parameters	(0, -400) (MWh)

Table 6: Experimental settings used to generate the time series in the electricity market problem.

Tests	Details
Unit tests	They verified the correctness of the log-normal distribution sampling and the calculation of the eutrophication threshold.
Property-based test	It verified whether lake eutrophication is irreversible in the absence of regulatory intervention.
Scenario test	It verified whether the program’s performance was consistent with the standard implementation from EMA Workbench [ Kwakkel, 2017 ] in a specific scenario (described in Section 4.3.3 ).

Table 7: Tests used for verification in Python implementation in the lake problem.

Tests	Details
Unit tests	They verified that the profit and penalty calculations for the wind-power producer were correct and that total offered supply was sufficient to meet the market demand.
Property-based test	It verified the correctness of the implementation of the merit-order market-clearing mechanism.
Scenario test	It verified whether the program’s performance was consistent with the baseline implementation in a specific scenario (described in Section 4.3.3 ).

Table 8: Tests used for verification in Python implementation in the electricity market problem.

	State variables	Decision variables	Transition function	Stochastic variables	Constants	Deep uncertainties	Objective function	Other component
Lake (before)	1	1	1	1	1	5	1	0.25
Lake (after)	1	1	1	1	1	5	1	1
Comprehensive market (before)	0	2	1	6	9	8	1	1
Brief market (before)	0	2	1	5	0	0	0	0
Market (after, for both)	0	2	1	6	10	8	1	1

Table 9: Number of components correctly extracted by the LLM before and after refinements in Step 1. These numbers were averaged over four runs in each sub-case. Results for Step 2-4 were not reported because all these components were correctly identified by the end of Step 1, and were preserved with minor errors in subsequent steps. The two sub-cases of the electricity market problem share the same “after” results because they are based on the same baseline description.

	Lake Problem [mean $\pm$ std]	Electricity Market Problem (comprehensive) [mean $\pm$ std]	Electricity Market Problem (brief) [mean $\pm$ std]
Initial formalization	$0.8\pm 0.4$	$0.8\pm 0.4$	$4.5\pm 0.5$
Multi-perspective formalization	$0.3\pm 0.4$	$0.5\pm 0.5$	$0.3\pm 0.4$
Composition	$0.0\pm 0.0$	$0.0\pm 0.0$	$0.0\pm 0.0$
Python implementation	$2.5\pm 0.5$	$1.5\pm 0.9$	$1.3\pm 0.4$

Table 10: Average number of iterations required by the LLM to produce a consistent response at each step in each sub-case. These numbers were averaged over four runs.

Panel of Figure 4 — Figure 4: Average time series of lake pollution for both perspectives, with a fixed sequence of pollution emission $\bar{a_{t}}$ (decisions from Perspective 1: local community), and varying pollution removal levels $r_{t}$ (decisions from Perspective 2: environmental regulator). To generate each of these figures, we ran the Python implementation produced in each experimental run ten times, collecting a total of 40 time series. Figure 4(a) shows the impact of Perspective 1 on the system alone, while Figure 4(b) – 4(d) show the combined impact of Perspective 1 and the increasing influence of Perspective 2 on the system.

Panel of Figure 5 — Figure 5: Average time series of clearing prices, actual dispatched wind energy and wind-power revenue in both sub-cases of the electricity market problem for both perspectives, with a fixed sequence of bids for the wind-power producer $(\bar{b_{wt}},\bar{p_{wt}})=(300,50)$ (decisions from Perspective 1: wind-power producer), and varying shortfall penalty coefficients $q_{u}$ (decisions from Perspective 2: system regulator). To generate each of these figures, we ran the Python implementation produced in each experimental run ten times, collecting a total of 80 time series. Figure 5(a) - 5(c) show the impact of Perspective 1 on the system alone, while Figure 5(d) shows the combined impact of Perspective 1 and the increasing influence of Perspective 2 on the system.

Exploring Plan Space through Conversation: An Agentic Framework for LLM-Mediated Explanations in Planning
Mar 2, 2026Guilhem Fouilhé, Rebecca Eifler, Antonin Poché +2Multi-Agent LLM Systems Explainable Artificial Intelligence
IRIT, Toulouse, France · LAAS-CNRS, Toulouse, France · IRT Saint Exupery, Toulouse, France +2
Can AI Reason Like an Urban Planner? Benchmarking Large Language Models Against Professional Judgment
Jun 10, 2026Yijie Deng, He Zhu, Wen Wang +3LLM Evaluation Urban Planning
Behavioral and Spatial AI Lab, Tongji University · Behavioral and Spatial AI Lab, Peking University · College of Architecture and Urban Planning, Tongji University
LifePlanner: Evaluating LLM Agents for Geo-spatial Planning with Social Media Data
Aug 25, 2026Zhen Dong, Yuning Peng, Yutao Shi +4Geospatial Reasoning Benchmarks LLM Agent Evaluation
Wuhan University · Hong Kong University of Science and Technology

Mar 2, 2026J/K move · Enter open · S save

Exploring Plan Space through Conversation: An Agentic Framework for LLM-Mediated Explanations in Planning

Guilhem Fouilhé, Rebecca Eifler, Antonin Poché, Sylvie Thiébaux+1

IRIT, Toulouse, France · LAAS-CNRS, Toulouse, France · IRT Saint Exupery, Toulouse, France+2

Multi-Agent LLM Systems Explainable Artificial Intelligence Human-in-the-Loop AI

Open paper AskPDF arXiv

When automating plan generation for a real-world sequential decision problem, the goal is often not to replace the human planner, but to facilitate an iterative reasoning and elicitation process, where the human's role is to guide the AI planner according to their preferences and expertise. In this context, explanations that respond to users' questions are crucial to improve their understanding of potential solutions and increase their trust in the system. To enable natural interaction with such a system, we present a multi-agent Large Language Model (LLM) architecture that is agnostic to the explanation framework and enables user- and context-dependent interactive explanations. We also describe an instantiation of this framework for goal-conflict explanations, which we use to conduct a user study comparing the LLM-powered interaction with a baseline template-based explanation interface.

Abstract

Explore similar work

Exploring Plan Space through Conversation: An Agentic Framework for LLM-Mediated Explanations in Planning

Can AI Reason Like an Urban Planner? Benchmarking Large Language Models Against Professional Judgment

LifePlanner: Evaluating LLM Agents for Geo-spatial Planning with Social Media Data