Towards Human-Like Behavior Generation by Applying Markov Chain Monte Carlo in Realization of Agential Realism

As the bridge between intelligence and action, decision-making reflects how humans adaptively interact with complex world. Thus, modeling decision-making process have become a key component for AI agents to handle dynamic environment. In our previous work, human decision-making process was modeled on a Decision Map, in which human decision strategies (in terms of individual or a group) are characterized by a decision distribution. In this paper, we apply the personal Decision Map with Markov Chain Monte Carlo (MCMC) algorithm to generate action sequences as a stochastic process. Based on the generated sequence, a Markovian transition matrix is estimated as a dynamic model for artificial agents to generate behaviors. The proposed method is inspired by the theory of agential realism, where the phenomenon of lightning is illustrated and manifested that its path is not fully random, but indeterminate caused by electromagnetic factors in the environment. In this sense, the generation of action sequence is considered as a lightning path, and different Markovian transition matrices can configure the dynamic indeterminacy for generating agent behaviors.