This paper points out the limitations of existing neuroscience decision-making research, which focuses on simplified behavioral tasks and explicit rewards, and deals only with repetitive and fixed behaviors of animals. In natural environments, animals often exhibit complex behaviors over long periods of time due to unobservable internal motivations. Time-varying inverse reinforcement learning (IRL) has been used to capture this, but it fails to consider that animal decisions are based not only on the current state but also on past history. In this paper, we present SWIRL (SWitching IRL), a novel framework that integrates time-varying and past-dependent reward functions. SWIRL models long-term action sequences as transitions between short-term decision processes, each governed by a unique reward function, thereby capturing how past decisions and environmental context shape behavior. We apply SWIRL to simulated and real animal behavior datasets and demonstrate that it quantitatively and qualitatively outperforms models without past dependence. This is the first IRL model that integrates past-dependent policies and rewards, advancing our understanding of complex and natural animal decision-making.
