An agent-based framework for modelling linked decisions

An agent-based framework for modelling linked decisions Sascha Holzhauer1,2 [0000-0003-1654-7433] and Friedrich Krebs1,2 [0000-0002-4104-8751] 1 University of Kassel, Wilhelmshöher Allee 71-73, 34121 Kassel, Germany 2 Fraunhofer Institute for Energy Economics and Energy System Technology, Joseph-Beuys-Straße 8, 34117 Kassel, Germany [email protected] Keywords: Linked decisions, framework, agent-based modelling, investments. 1 Motivation The energy transition towards CO2-neutrality requires substantial financial investments. Large-scale infrastructure investments must be complemented by small-scale private investments for instance regarding private building retrofitting or electric mobility. The small-scale private investors differ in at least two respects to professional investors. First, they are laypersons regarding financial investments in general and with respect to their domain knowledge on the technical properties of the investment good. Second, they are confronted with multiple investment domains in parallel [1]. That is, investments in several goods are to be considered at the same time, and it is not defined which investment decision is made first, if any. In particular, the latter property adds significant complexity to the decision situation because numerous factors impact the planning of an implementation sequence of technology investments, such as available capital, individual preferences, current state of a building, funding, or neighbouring activities. In the described problem domain governed by interaction, dynamics, bounded rationality, and actor heterogeneity, agent-based simulation can contribute substantial insight [2]. A flexible yet well-founded decision-making concept which allows to model the outlined parallel, multi-domain investment decisions of small-scale private investors is still missing [3]. However, the nature of linked decisions has been described in the field of organisational decision making, and the considered situation can be labelled a lateral, pooled linkage because of shared resources [4]. Due to the lack of further psychological and economic literature on the underlying decision processes, we base our model concept on assumptions: • Decision triggers: Investment decisions are triggered by events external to the actor. The consideration of an initial decision may trigger further decisions to be made concurrently. • Avoiding path dependencies of decision sequences: Decisions may not be treated in a pre-defined row without limitations because the outcome of one decision has potential implications on further decisions (e.g. initially investing in a costly option impedes investments in cheaper but more desired options of other decision domains). 2 • Representation of technical interdependencies: The final order of decision implementations is important to model technical interdependences between decisions to target interventions (e.g. incentivise insulation before exchange of heating system). In the following, we describe the model concept and implementation. Following this, we illustrate the decision model regarding the case of building retrofitting including investments in heating, insulation, windows, rooftop photovoltaic (PV), electric vehicle (EV) and discuss the approach. 2 Approach Considering the required resources in terms of both finances as well as cognitive effort to plan and implement, and estimating the consequences of the investments in terms of property value, maintenance, and comfort, deliberative reasoning seems plausible. However, examining a likely situation when house owners are triggered to exchange their heating system and consider insulation measures, installation of PV and purchase of an electric car, assuming only two options plus the non-investment option for each single decision, this sums up to 3^4 = 81 combinations of decision options. Likely, this is too much to evaluate for a couple of attributes (e.g. investment costs, environmental impact, comfort, installation effort, availability) and to decide rationally. Fig. 1. Flow chart of the linked decision-making process. Boxes framed in red are subject to custom implementations (yellow marks the decision domain selection, blue the choice of options). Text in white refers to the “wheel of fortune” decision selection variant. As an alternative, decision makers may consider investments sequentially. However, the nature of interdependent investments in combination with limited budgets hinders a defined order, which anyway strongly depends on soft factors such as exposure to 3 certain information sources, personal preferences and availability of installers. Consequently, the order of investments can hardly be predefined and – once defined – cannot guarantee successful execution. Therefore, we propose an iterative approach for linked multi-investment decisions within a flexible framework. The principal idea is to select an investment decision with a previously defined probability, gather options, make the decision, and update remaining resources for further investments. The subprocess is repeated until all triggered decisions are made or any resource is depleted. Finally, the common set of chosen options is checked for viability, and the process is repeated in case it is not. The approach offers a high degree of freedom and may account for various contexts, as there are placeholders for flexible implementations. We discuss the flexibility while describing in more detail the algorithm shown in Fig. 1. Often, house owners are exposed to a specific event, which then triggers several investment decisions. For instance, new funding opportunities for electric cars trigger decisions about PV installation. Similarly, a breakdown of the heating system triggers decisions about insulation and PV installation. We refer to the particular kind of investment as decision domain. Initially, all the relevant decision domains are collected, and data to derive the order of execution is gathered. As default implementation, each decision domain is assigned a probability of being selected on the “wheel of fortune”. Here, individual preferences play a key role, as some would select investments first that are estimated to be most expensive, others consider the urgency of decisions strongly. It is furthermore possible to represent investments with a defined order as a combination of decisions. For instance, the decision about a wall box usually follows the decision about an electric car because its type depends on the particular EV. After the “wheel of fortune” is set up with probabilities it can be turned and a decision is selected to be made, which triggers the according decision-making process. As first part of that process, possible options are gathered. The algorithm which determines the chosen option can be defined per decision, i.e. for a wall box a heuristic such as imitation can be applied while the decision for a new heating system can be based on rational evaluations of all available options. Once a decision is made, the chosen option is added to the set of all chosen options, and the resources this option would consume are subtracted from the remaining resources. The “wheel of fortune” is adapted (without the already selected decision domain) and turned again. When all decision domains have been selected or one of the considered resources is exhausted, the algorithm suggests a check of feasibility to implement the entire set of selected investment options. For example, in case of a broken heating system the investments should contain a new heating system and not exclude it because of running out of budget. If the check fails, optionally the probabilities for decision domains can be recalculated and the selection of decision starts again. 4 3 Implementation The afore-mentioned flexibility is not only reflected by the liberal calculation of probabilities for decision domains and the set of considered resources. Four layers of implementation classes specify parts of the algorithm and can be substituted by appropriate versions (in Fig. 2). The DecisionSelector is responsible for selecting the decision domain. Whereas the “wheel of fortune” approach acts as default, alternative implementations may e.g. strictly select the decision with highest probability. The DecisionDomain is responsible to gather possible options and can be used to filter these. Most important, it specifies the way of decision-making by providing the DecisionDecider. Which decider is provided may depend e.g. on the agent type. The decider can be anything from the application of heuristics such as neighbourhood imitation to completely rationally choosing the best performing option regarding all attributes. The framework does not restrict applicable models of agents’ decision making (see [5] for an overview and [6] on guidelines for selection and justification). DecisionOptions support the DecisionDecider in evaluation themselves, and implement their execution, e.g. installing the heating system they represent. Fig. 2. The four layers of implementing the linked decision-making framework. 4 Discussion We present a flexible framework for modelling linked investment decisions. The probability-based selection of decision addresses unspecified soft factors but also allows prioritisation on individual or environmental attributes. Though designed for the application to linked energy-related property investments, the approach can be applied to any kind of interdependent decisions without a clearly defined order. Further examples are medical treatments or spending leisure time. In future, we seek to strengthen the empirical foundation for the flexible parts of the framework and according parameters following the KIDS approach. As available theory on linked investment decisions is insufficient, we call for dedicated studies on this subject. Elaborated sensitivity analyses are planned to inform such studies. 5 References 1. Gu, G.; Yang, D.; Feng, T. & Timmermans, H. Influence of the adoption of new mobility tools on investments in home renewable energy equipment: Results of a stated choice experiment. Sustainable Cities and Society, Elsevier BV, 2019, 50, doi:10.1016/j.scs.2019.101641 2. Hesselink, L. X. & Chappin, E. J. Adoption of energy efficient technologies by households – Barriers, policies and agent-based modelling studies. Renewable and Sustainable Energy Reviews, Elsevier BV, 2019, 99, 29-41. doi:10.1016/j.rser.2018.09.031 3. Senkpiel, C.; Dobbins, A.; Kockel, C.; Steinbach, J.; Fahl, U.; Wille, F.; Globisch, J.; Wassermann, S.; Droste-Franke, B.; Hauser, W.; Hofer, C.; Nolting, L. & Bernath, C. Integrating Methods and Empirical Findings from Social and Behavioural Sciences into Energy System Models—Motivation and Possible Approaches. Energies, MDPI, 2020, 13, 4951, doi:10.3390/en13184951 4. Langley, A.; Mintzberg, H.; Pitcher, P.; Posada, E. & Saint-Macary, J. Opening up Decision Making: The View from the Black Stool. Organization Science, Institute for Operations Research and the Management Sciences (INFORMS), 1995, 6, 260-279, doi: 10.1287/orsc.6.3.260 5. Balke, T. & Gilbert, N. How Do Agents Make Decisions? A Survey. Journal of Artificial Societies and Social Simulation, Journal of Artificial Societies and Social Simulation, 2014, 17, doi: 10.18564/jasss.2687 6. Wijermans, N.; Scholz, G.; Chappin, É.; Heppenstall, A.; Filatova, T.; Polhill, J. G.; Semeniuk, C. & Stöppler, F. Agent decision-making: The Elephant in the Room - Enabling the justification of decision model fit in social-ecological models. Environmental Modelling & Software, Elsevier BV, 2023, 170, 105850, doi: 10.1016/j.envsoft.2023.105850