# Adaptive Agents in Spatial Double-Auction Markets: Modeling the Emergence of Industrial Symbiosis

Crypto.world

> Industrial symbiosis—one firm's waste becoming another firm's input—promises environmental and economic gains, but exchanges rarely appear on their own because transport costs, supply-demand…

**arXiv**: 2512.17979

**Published**: 2025-12-19

**Authors**: Matthieu Mastio, Paul Saves, Benoit Gaudou, Nicolas Verstaevel

## AI summary

## Overview

**Research area:** Multi-agent systems and agent-based simulation applied to market design, industrial symbiosis, and the circular economy (arXiv category cs.GT; presented at AAMAS 2026).

**Technical level:** Advanced. The paper assumes familiarity with double auctions, multi-agent reinforcement learning (bandit methods, Boltzmann exploration), counterfactual regret, and variance-based global sensitivity analysis with surrogate models (Polynomial Chaos Expansion, Sobol' indices, PDP/ICE plots).

**Scope (one sentence):** The paper builds and simulates a spatially embedded, decentralized double-auction market in which firms learn byproduct pricing strategies through reinforcement learning, in order to identify the spatial and economic conditions under which industrial symbiosis emerges.

## What This Paper Is About

Industrial symbiosis—one firm's waste becoming another firm's input—promises environmental and economic gains, but exchanges rarely appear on their own because transport costs, supply-demand mismatches, and coordination failures get in the way. Existing agent-based symbiosis models mostly give firms fixed behavioral rules or static prices, and existing auction/RL market models mostly ignore geography, so neither captures how adaptive, spatially situated firms actually converge on trades. The authors' goal is to fill that gap: model heterogeneous buyers and sellers trading one byproduct through a decentralized multilateral auction on a simulated territory, let sellers learn their pricing markups over time, and then measure when prices stabilize and how much local circularity results.

## Key Contributions

1. **A decentralized, spatially explicit multilateral double-auction mechanism** for byproduct exchange in which transaction prices *and* traded quantities emerge endogenously from local interactions rather than being set exogenously. Seller prices take the form `p_ij = φ_j · p_m + c_ij`, where `c_ij = d_ij · c_t` internalizes distance-based transport cost.
2. **Learning sellers:** each seller's pricing decision is framed as a contextual multi-armed bandit over a discretized action space of `K` markup intervals, with reward estimates updated by an exponential moving average and actions selected by Boltzmann (softmax) sampling with a decaying temperature.
3. **Validation via counterfactual regret analysis**, showing that per-step regret declines toward zero and the joint strategy profile approaches an approximate Nash equilibrium—evidence that the emergent market dynamics are stable, not merely simulated noise.
4. **A global variance-based sensitivity analysis** (sparse Polynomial Chaos Expansion and a neural-network surrogate, yielding Sobol' indices plus Partial Dependence and Individual Conditional Expectation plots) that ranks which spatial and economic parameters drive exchange price and the newly defined symbiosis index, separating actionable policy levers from structural constraints.

## Main Findings

- **Price convergence depends on scarcity and disposal cost:** With density fixed at 0.001, high scarcity (`s = 2`) and low disposal cost (`c_d = 10`) drove mean transaction prices toward an equilibrium close to the external market price. Low scarcity (`s = 0.5`) with `c_d = 20` drove prices steadily down to stabilize near the disposal cost, reflecting buyers' stronger bargaining power.
- **Strategies approach a near Nash equilibrium:** Rolling-median (window = 50) per-step counterfactual regret declined over time and converged toward zero in both scarcity/disposal scenarios, meaning sellers were close to the best fixed action available in hindsight at each step.
- **Disposal penalties pull prices down, with scarcity as moderator:** Higher `c_d` induced lower prices, most strongly at `s = 0.5`, where in dense territories equilibrium prices could fall below zero. At `s = 1` the downward trend was attenuated; at `s = 2` prices were largely insensitive to `c_d` because excess demand kept buyers paying near the outside option `p_m`. Variance was highest at `s = 1`, indicating oscillating strategies under balanced competition.
- **Symbiosis rises with disposal cost but is capped by density:** The symbiosis index `SI = q_bought / min(q_toSell, q_needed)` increased with `c_d`; in dense environments local exchange saturated close to 1, while in sparse environments symbiosis stayed low even under high `c_d`. Scarcity had only a marginal effect on symbiosis compared with disposal cost and density.
- **Firm density dominates both outcomes:** In the Sobol' analysis, density `ρ` explained `S_1 = 63.2%` of variance in symbiosis and `20.5%` in price—the largest single share in each case. Transport cost `c_t` contributed `S_1 = 16.7%` (symbiosis) and `8.5%` (price); scarcity `s` contributed `S_1 = 24.6%` to price but only `1.3%` to symbiosis; disposal cost `c_d` contributed `11.0%` to price and `2.4%` to symbiosis. Higher-order interactions were substantial (e.g., `s,ρ` at `S_2 = 14.0%` for price; `ρ,c_t` at `S_2 = 7.5%` for symbiosis). Unaccounted variance was ≈4.7% for price and ≈3.0% for symbiosis.
- **Scarcity shows a threshold effect on price:** In the PDP/ICE analysis, scarcity had little effect at low levels but strongly pushed prices upward between 1.5 and 2.5; beyond 2.5 the effect persisted but grew more slowly.
- **Cluster spread is a weak and conditional lever:** It had limited influence on price overall, and mattered for symbiosis only at low density and only at small spread values—spatial compactness can partially compensate for low density, but its influence fades rapidly as spread increases.
- **Surrogate models were accurate:** The sparse PCE reached `R²` of 95% (symbiosis) and 92% (price) on a 20% test sample; the MLP surrogate reached `R²` of 97% (symbiosis) and 95% (price).

## Methodology in Plain English

The authors built a simulated territory populated by two kinds of firms: buyers who need a byproduct and sellers who hold it as waste. Each firm sits at a location, so every potential trade carries a transport bill proportional to distance times a per-kilometer cost. Buyers differ in how much they need and in how price-sensitive they are—a buyer only accepts an offer at or below its own tolerance relative to an external market price, which is treated as fixed because the local park is assumed too small to move global prices.

Each round works as a negotiation round: every seller sends every buyer a personalized offer (a markup on the market price plus the distance cost to that buyer), each buyer picks its favorite acceptable offer, and each seller then sorts the acceptances by profit per unit and confirms contracts until it runs out of stock. Satisfied buyers and emptied sellers drop out and the round repeats until nothing more can be matched. Sellers that end a round with unsold stock pay a landfill penalty, which enters their reward.

Sellers improve over time by trial and error. Their only decision is which markup to charge, chosen from 30 discrete options. They keep a running score for each option, updated as a moving average of the rewards it earned, and pick options probabilistically with a "temperature" that starts high (explore) and cools (exploit). The authors report that this softmax scheme converged faster with less oscillation than epsilon-greedy or UCB alternatives in their multi-agent setting.

To check whether the learning actually worked, they computed, at each step, how much more a seller *could* have earned by picking its best alternative given what everyone else did, and showed this gap shrinks to near zero. To see which knobs matter, they ran 20,000 simulations in total, fit fast surrogate models to the simulator, and used those surrogates to attribute the variance in final price and in the symbiosis index to five inputs: disposal cost, scarcity, firm density, cluster spread, and transport cost per kilometer.

**Experimental settings reported:** 40 firms (described as the typical scale of a standard EIP), four spatial clusters, default cluster spread of 1 (near-uniform), market price fixed at 100, transport cost 0.1 per kilometer, temperature decay 0.996, 30 action bins, 1000 timesteps per run, and 10 runs averaged per point in the price/symbiosis figures. Simulations used virtual 2D environments rather than a real road network; the reported throughput was roughly 100 simulations of 1000 timesteps per minute on an Intel Core Ultra 9 185H (21 cores). Code is open source.

**Not reported in the supplied content:** the paper's Discussion (Section 6) and Conclusions (Section 7) are cut off in the provided text, so any discussion points or concluding claims are not summarized here. The truncation also ends Section 5.3 mid-sentence, partway through the symbiosis PDP/ICE interpretation.

## Why This Matters

**Research impact:** The paper closes three gaps it explicitly identifies—industrial symbiosis models lacking endogenous market-based price formation, auction/negotiation models lacking adaptive learning, and reinforcement learning market models lacking decentralized spatial topology. It shows that a relatively lightweight bandit formulation, not deep multi-agent RL, is enough to produce converging, near-equilibrium behavior in a spatial market, and it demonstrates a reusable pipeline (simulator → surrogate → Sobol'/PDP) for ranking policy levers in agent-based models.

**Real-world applications:**

- **Eco-industrial parks (EIPs):** the stated target setting, where park managers need to know whether to invest in clustering firms or in raising landfill penalties.
- **Heat recovery networks** between industrial and urban actors, where distance-based cost is the dominant friction.
- **Wastewater reuse and biomass valorization**, both named by the authors as transferable circular-economy settings.
- **Aggregated municipal or neighborhood actors** that must coordinate resource loops under transport friction, rather than only individual firms.

**Industry relevance:** The headline policy lesson is that disposal-cost penalties and firm density, not cluster compactness or per-kilometer transport cost, are the strongest levers—density alone accounts for 63.2% of the variance in symbiosis. That suggests subsidies for co-location or landfill pricing are more effective than fine-tuning logistics costs, and that sparse territories cannot be rescued by penalty pricing alone. The finding that prices can fall below zero under low scarcity and high disposal costs is directly relevant to firms deciding whether local exchange is worth participating in.

## Future Directions

- **Validate on real geographies:** the framework supports road networks and real territories, but all reported experiments used abstract 2D environments; testing on an actual park or region is the obvious next step.
- **Extend beyond a single byproduct:** the authors deliberately restrict to one product "to concentrate on essential dynamics," leaving multi-byproduct, interdependent exchange—where one firm's input is another's waste in a chain—unexplored.
- **Test policy interventions directly:** the paper positions the model as "a basis for exploring policy interventions that seek to align firm incentives with sustainability goals," but no specific intervention (subsidy, tax, matching platform) is simulated in the reported results.
- **Address the sparse-territory problem:** since symbiosis stayed low in sparse environments even under high disposal costs, a natural open question is what complementary mechanism—brokering, aggregation, or shared logistics—could raise circularity where density cannot be increased.

## Target Audience

Researchers in multi-agent systems and agent-based modeling, especially those working on market design, auctions, and multi-agent reinforcement learning; industrial ecologists and circular-economy modelers who want endogenous prices instead of fixed behavioral rules; and policy analysts or EIP planners interested in which spatial and economic levers most affect byproduct exchange. Readers should be comfortable with auction mechanics and RL terminology, though the plain-language framing of the model makes the results section accessible to economists and sustainability researchers without an RL background.

## Authors’ abstract

Industrial symbiosis fosters circularity by enabling firms to repurpose residual resources, yet its emergence is constrained by socio-spatial frictions that shape costs, matching opportunities, and market efficiency. Existing models often overlook the interaction between spatial structure, market design, and adaptive firm behavior, limiting our understanding of where and how symbiosis arises. We develop an agent-based model where heterogeneous firms trade byproducts through…

[The full abstract on arXiv](https://arxiv.org/abs/2512.17979)

[Read the original paper](https://arxiv.org/abs/2512.17979)

---

Canonical: https://crypto.world/paper/adaptive-agents-in-spatial-double-auction-markets-modeling-the-2512.17979
