Abstract:
Agent-Based Models (ABMs) have become more widespread over the last two decades allowing researchers to explore complex systems composed of heterogeneous entities. Although ABMs have proven effective for generating simple rules over homogenous and heterogeneous agent types to observe emergent behaviors, several challenges exist. One, typical ABMs are limited in the representation of cognition and learning to maximize their actions based on current (and future) rewards of being in a particular state. Two, ABMs are not designed to produce their own behaviors that can be interpreted by the designer. Although agents may act upon code generated by the model designer, their local and global responses are not easily interpretable. Additionally, ABMs do not decompose behaviors into information rooted in cognitive processing, specifically satisficing or fast-and-frugal heuristics. To address these challenges, this dissertation presents a model and methodology called the Learning-based Actor-Interpreter State Representation (LAISR), where agents use Deep Reinforcement Learning (DRL) to generate strategies to maximize for current and future states. Due to the agent’s behavior representation as a deep neural network (DNN), explainable artificial interpretation (XAI) methods are used to decompose DNN features into simple but satisficing strategies (heuristics). The results of this work demonstrate an approach that bridges machine learning with that of the social sciences where agents can build their own optimal and boundedly rational strategies. This methodology is demonstrated across several homogeneous and heterogeneous agent-based models. The implications of this work demonstrate significant steps towards how machine learning-enhanced ABMs can be used to develop novel and optimal decision strategies, enhance human behavior modeling, and provide a bridge between social science and artificial intelligence research.
