{"paper_id":"2501.09429","title":"ADAGE: A generic two-layer framework for adaptive agent based modelling","authors":["Benjamin Patrick Evans","Sihan Zeng","Sumitra Ganesh","Leo Ardon"],"paper_date":"2025-01-16","paper_type":"perspective-policy","paper_url":"https://arxiv.org/pdf/2501.09429v1","page_url":"https://thequant.space/flowcharts/adage-a-generic-two-layer-framework-for-adaptive-agent-base/","scores":{"empirical_rigor":6,"math_complexity":8.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes ADAGE, a generic two‑layer Stackelberg‑game framework that jointly learns adaptive agent behavioural policies and environmental changes.","quote":"In this work, we develop a generic two-layer framework for ADaptive AGEnt based modelling (ADAGE) for addressing these problems. This framework formalises the bi-level problem as a Stackelberg game with conditional behavioural policies, providing a consolidated framework for adaptive agent-based modelling based on solving a coupled set of non-linear equations.","summary":"Agent-based models (ABMs) are valuable for modelling complex, potentially out-of-equilibria scenarios. However, ABMs have long suffered from the Lucas critique, stating that agent behaviour should adapt to environmental changes. Furthermore, the environment itself often adapts to these behavioural…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"The bi‑level adaptive problem can be formalised as a Stackelberg game with a leader (outer layer) and followers (inner simulation layer).","type":"model","stated":"explicit","quote":"Therefore, in the following section, we formulate adaptive agent-based model design as a Stackelberg game, with the leader as the outer layer, and the followers as the (inner) agent-based simulation layer. 3 PROPOSED APPROACH We propose a generic two-layer framework, ADAGE, as visualised in fig. 1.","support":"full"},{"assumption":"Followers condition their behavioural policies on observed environment characteristics (θ̂).","type":"model","stated":"explicit","quote":"In the rest of the paper, we denote the behavioural policy of agent 𝑖by 𝜋𝑖(𝑎| 𝑜𝑖, ˆ𝜽𝑖) to emphasise that the followers condition their behavioural policies on the (observed) environment characteristics. When the conditioning is not explicit, it is assumed that ˆ𝜽𝑖is part of the observation 𝑜𝑖.","support":"full"},{"assumption":"The gradient domination condition holds for the agents' return functions, making every stationary point globally optimal.","type":"statistical","stated":"explicit","quote":"Due to the “gradient domination” condition satisfied by eq. (2), every stationary point is globally optimal [1].","support":"full"},{"assumption":"When the number of followers is large, shared policy learning can reduce the number of distinct follower policies.","type":"model","stated":"explicit","quote":"(4) Note that in cases of large 𝑛, in order to reduce the number of equa-tions that must be solved, shared policy learning [52] can be used for agents of similar types, reducing the set of follower behavioural policies to a subset (<< 𝑛) of shared behavioural policies.","support":"full"},{"assumption":"The outer‑layer learning rate (αL) is chosen much larger than the inner‑layer rates (αF) to approximate a nested‑loop algorithm.","type":"execution","stated":"explicit","quote":"In general, we need to choose 𝛼𝐿,𝑡to be much larger than all 𝛼F,𝑡to approximate a nested-loop algorithm that runs multiple inner layer updates per outer layer update.","support":"full"},{"assumption":"Information processing costs for boundedly rational agents are modelled via a KL‑penalty term.","type":"model","stated":"explicit","quote":"While different information cost functions 𝐼can be used, here we use the KL penalty [17, 36]: 𝐼(𝜋,𝑜) = ∑︁ 𝑎 𝜋(𝑎|𝑜) log 𝜋(𝑎|𝑜) 𝜋0(𝑎) where 𝜋0(𝑎) is some prior belief, which we assume to be uniform.","support":"full"}]},"datasets":{"status":"available","items":[{"name":"TaxAI","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"To demonstrate the use of ADAGE for policy design, we base an environment on TaxAI, \"the most realistic eco-nomic simulator for optimal tax policy\" [33], based on the canonical Bewley-Aiyagari model. Under this environment, there are multi-ple households, firms, banks, and a central government (visualised in fig. 2).","support":"full"},{"name":"Cobweb market game","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"To demonstrate how calibration can be per-formed with ADAGE, we use the Cobweb market game [27] to explore market price fluctuations, where with human participants, we see larger fluctuations than is to be expected by perfectly ratio-nal participants, so we must calibrate the information processing costs (modulating the bounded rationality) of the agents [16]. In a cobweb market, 𝑖∈F producers must estimate the","support":"full"},{"name":"Financial market maker scenario","provider":null,"kind":"synthetic","access":"unknown","period":null,"universe":null,"quote":"We demonstrate robust behavioural learning under a financial environment with a market maker (MM) trading with various Liquidity Takers (LTs), where the MM may have differ-ent preferences in terms of maximising market share or profit and loss (PnL), and the goal is to learn generic MM behaviour across these preferences. There is one learning MM agent and 𝑛 zero-intelligence LT agents [18, 30] interacting in a market.","support":"full"}],"known":[],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"synthetic"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R1","candidates":{"front_matter":"n/a","evidence_card":"no","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"unclear","findings":[{"finding":"ADAGE improves social welfare in the TaxAI policy‑design experiment, with the mean welfare increasing over training iterations.","kind":"performance","quote":"0 200 400 600 800 1000 Training Iterations Low 8 6 4 2 0 High Social Welfare ADAGE Free Market Figure 3: Social welfare throughout training.","support":"full"},{"finding":"Calibration of the Cobweb market using ADAGE reduces mean absolute error (MAE) to 0.005 and RMSE to 0.007, outperforming the base simulator.","kind":"other","quote":"Table 1: Calibration metrics (based on the mean error from multiple rollouts to bootstrapped samples from experimen-tal data of [27]). Equilibria Base Simulator ADAGE MAE 0.01 0.013 0.005 RMSE 0.017 0.026 0.007 Producers receive a non-linear reward based on the accuracy of their prediction compared to the resulting market price 𝑝𝑡[27]: 𝑟F,𝑡= max(0, 1300 −260(𝑝𝑡−ˆ𝑝𝑖,𝑡)2) (12) Action The action for the calibrator is a","support":"full"},{"finding":"In the scenario‑generation task, the ADAGE‑derived Tobin’s tax reduces market volatility in the majority of configurations (7/9 cases for std dev, 6/9 for MAPC).","kind":"performance","quote":"The outer layer discovered automatically that scenarios with a Tobins tax successfully restricts the market volatility (reduces market fluc-tuations) in the vast majority of configurations. This reduction is observed in the standard deviation of the attendance time series (7/9 cases) and the mean absolute percentage change (6/9 cases).","support":"full"},{"finding":"Robust behavioural learning enables a market‑maker to adapt across differing preferences without loss of performance, demonstrating algorithm‑independence.","kind":"theoretical","quote":"This shows an additional benefit of ADAGE: when specific behaviour can be computed analytically, this can be substituted directly for any of the behavioural policies, and the remaining behaviour will be optimised w.r.t this, again demonstrating independence on the learning algorithm.","support":"full"},{"finding":"ADAGE can substitute analytically derived policies for learned ones when available, preserving the optimisation of remaining behaviours.","kind":"theoretical","quote":"This shows an additional benefit of ADAGE: when specific behaviour can be computed analytically, this can be substituted directly for any of the behavioural policies, and the remaining behaviour will be optimised w.r.t this, again demonstrating independence on the learning algorithm.","support":"full"}],"reported_performance":null,"out_of_sample":{"method":"none","headline_out_of_sample":null}},"independent_evidence":{"status":"not_extracted","supports":[],"contradicts":[],"qualifies":[],"compared_with":0,"citations":{"cited_by_in_corpus":0,"by_independent_authors":0,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":0},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":2,"extracted_at":"2026-11-14T19:53:04Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}