A methodical framework for engineering co-evolution for simulating socio-economic game playing agents
Agent based computational economics (ACE), as a research field, has been using co-evolutionary algorithms for modelling the socio-economic learning and adaptation process of players within games that model socio-economic interactions. In addition, it has also been using these algorithms for optimising towards the game equilibria via socio-economic learning. However, the field has been diverging from evolutionary computation, specifically co-evolutionary algorithm design research. It is common practice in ACE to explain the process and outcomes of such co-evolutionary simulations in socio-economic terms. However, co-evolutionary algorithms are known to have unexpected dynamics that lead to unexpected outcomes. This has often lead to mis-interpretations of the process and outcomes in socio-economic terms, a case in point being the lack of a methodical use of the term bounded rationality. This mis-interpretation can be attributed to the lack of a proper consideration of the solution concept being implemented by the co-evolutionary algorithm used for the simulation.
We propose a holistic methodical framework for analysing and designing co-evolutionary simulations, such that mis-interpretations of socio-economic phenomena be methodically avoided, disabling the algorithm from being mis-interpreted in socio-economic terms, aimed at benefiting ACE as a research field. More specifically, we consider the methodical treatment of co-evolutionary algorithms, as enabled by the framework, such that mis-interpretations of bounded rationality be avoided when these algorithms are used to optimise towards equilibrium solutions in bargaining games. The framework can be broken down into two parts:
1. Analysing and refining co-evolution for ACE, using the notion behind co-evolutionary solution concepts from co-evolutionary algorithm design research: Challenging the value of the implicit assumption of bounded rationality within co-evolutionary simulations, which leads to it being mis-interpreted, we show that convergence to the equilibrium solutions can be achieved with boundedly rational agents by working on the elements of the implemented co-evolutionary solution concept, as opposed to previous studies where bounded rationality was seen as the cause for deviations from equilibrium. Analysis and refinements guided by the presence of top-down equilibrium solutions, allow for a top-down avoidance of mis-interpretations of bounded rationality within simulations.
2. Analysing and refining co-evolution for ACE, using the notion behind reconciliation variables proposed in the thesis: Reasonably associating mis-interpreted socio-economic phenomena of interest with the elements of the implemented co-evolutionary solution concept, parametrising and quantifying the elements, we obtain our reconciliation variables. Systematically analysing the simulation for its relationship with the reconciliation variables or for its closeness to desired behaviour, using this parametrisation, is the suggested idea. Bounded rationality is taken as a reconciliation variable, reasonably associated with agent strategies, parametrised and quantified, and analysis of simulations with respect to this variable carried out. Analysis and refinements based on such an explicit expression of bounded rationality, as opposed to the erstwhile implicit assumption, allow for a bottom-up avoidance of mis-interpretations of bounded rationality within simulations.
We thus remove the causes that lead to bounded rationality being mis-interpreted altogether using this framework. We see this framework as one next step in ACE socio-economic learning simulation research, which must not be overlooked.
Available for download here.