This paper proposes a new state space approach to adaptive artificial intelligence (AI) modeling under the dynamic environment, where Bayesian filtering sequentially learns the model parameters including model structures themselves as state variables. In particular, our approach is widely applicable to the machine learning of non-linear AI models for real-time observation data flows through Monte-Carlo simulation-based filtering algorithms called particle filters.
To show the effectiveness of our framework, we concretely design a Takagi-Sugeno-Kang fuzzy model for financial portfolio construction, where particle filtering learns the model parameters as state variables. As a promising application, we suppose that the model parameters follow mean-reversion processes, which makes it possible to update these parameters around predetermined levels. Therefore, by deciding the levels based on existing state-of-art learning methods over the training data, our approach successfully incorporates and extends their learning results through adjusting those to the changing environment. An out-of-sample simulation with long term time-series data of stock and bond prices demonstrates the validity of our framework.