We investigate a heterogeneous network (HetNet) including sub-6GHz base stations (BSs), mmWave BSs, and THz BSs to support enhanced mobile broadband (eMBB) users and ultra-reliable low-latency communication (URLLC) users. We particularly investigate a user-centric network in which the users locally and dynamically select and switch among BSs over time to achieve their highest utility. Two types of users have different Quality of Service (QoS) requirements. Thus, we design two types of utility functions specifically for the eMBB users and URLLC users. Then, to model the dynamic selection behavior of the users, we propose to use a fractional game with the power-law memory. The fractional game allows the eMBB users and the URLLC users to incorporate their past strategies into their current selection, thus improving their utility. Furthermore, we consider the case that the BSs communicate the system state with each other, and we model the network selection of the users as a multi-agent problem. Then, we propose to use a multi-agent deep reinforcement learning (MADRL) algorithm that enables the URLLC users and eMBB users to make their network selection decision online to achieve their long-term utility. Various simulation results are provided to demonstrate the scalability and effectiveness of the proposed approaches. Particularly, compared with the classical game, the fractional game is able to achieve a higher utility but incurs a higher network adaptation cost. Moreover, the different types of URLLC users (in terms of latency and reliability requirements) and the number of URLLC users in the network significantly affect the total utility and the network selection strategies of the eMBB users. Importantly, given the full observations, the MADRL outperforms both classical and fractional games in terms of total network utility.