Strategic Earning on Tokenized Platforms via Model-based Decision Making

We develop a decision-making framework for determining participant's investment and laboring intensity on tokenized platforms, i.e., digital platforms with blockchain implementation. The framework builds on the platform development model that integrates participants' three roles on tokenized platforms - user, investor, and laborer - in platform's multi-facet utility for participants. We extend the model to investigate strategic decisions on participants' investment and laboring intensity. Arbitrary participants are distinguished from the platform-average participant, and the decision-making is cast into two sub-problems: when ignoring individual actions' impact on platform's state, strategies are substantiated by metrics characterizing model projection on platform's future development; when considering participants' decisions explicitly influencing the environment, the setting becomes an optimal control problem, which we capture with an Markov decision process (MDP) and solve via reinforcement learning (RL). The framework addresses three uncertainties in this decision-making: parameter uncertainty from model estimation, system uncertainty in model projection, and input uncertainty during individual-platform interaction. Monte-Carlo (MC) methods are used to address parameter uncertainty and system uncertainty, by initiating an accepted pool for parameter estimates and an ensemble for projection instances, from which metrics are computed to substantiate strategy design; input uncertainty is addressed with the MDP and the RL solvers. For the metric-based approach, we analyze two decision settings: fixed amount execution, and fixed increment execution. For the RL-based approach, we consider multiple environments under fixed increment execution. The framework is tested with ground-truth token price series. The platform development model achieves considerable quality in projection. A panel of metric-based strategies for investment and laboring intensity are constructed, and certain strategies outperform model-free strategies. Two deep RL agents (DQN, PPO) are implemented; they behave rationally on synthetic states and are shown to outperform metric-based and model-free strategies on ground truth