Modelagem de QoE para jogos em nuvem

Abstract

The number of users of traditional gaming platforms (consoles and computers) in January 2020 was 312 million when adding the users of Steam, PlayStation and Xbox Live. The cloud gaming platform GeForce Now had 10 million total users as of April 2021. In addition, market reports predict a growth rate of cloud gaming between 50-60% per year for the next ten years. This dissertation proposes a model for QoE in cloud games that uses network and cloud parameters. In addition, we separately consider the network effects for video streams and game commands. As an output of the model, we have the predicted evalueted QoE by the user. Such models can be used to help manage network provider’s assets or a cloud provider’s computing assets, aiming to improve the user experience. Other possible applications are: estimating the impact of network changes from other cloud gaming works, validating requirements of a new cloud gaming platform, and balancing load by a network provider according to QoE prediction. Our work is the first to use a QoE survey for cloud gaming and create a network-based QoE predictor. Since the existing models are based only on QoS, our model is relevant as it can be used in scenarios that are not possible with the previous models.We used a testbed with server and client on the same network, generating network degradation per match according to ITU document P809. Data were collected from 2020 matches with nine users. The data was used to train a regressor for two different models, one with agreement and the other without agreement. The without agreement model assumes that the network provider and the cloud gaming provider do not hareinformation, so the model only has access to network data. The model with agreement, on the other hand, assumes that there is an agreement and has more data, allowing for a deeper analysis. Both models were able to predict the evaluated QoE within an error range of plus or minus one on a seven-point MOS scale for the without agreement model and plus orminus 0.9 for the with agreement model. The hit rate of the model without agreement is 40% and that of the model with agreement is 50%. This value goes up to 90% considering evaluations with error of up to one unit as correct. In addition, we analyze aspects such as the creation of hierarchical models, generalization and explainability.