Over the past few years, the surge in VR (Virtual Reality) video traffic on networks has been remarkable. Nonetheless, a key challenge remains: ensuring a top-notch quality of experience (QoE) for VR video playback, especially when network bandwidth is limited. Prior studies have mainly focused on tile-based adaptive bitrate (ABR) streaming operating at the application layer on the server/client side to improve QoE, using single viewport prediction to conserve bandwidth. However, single-viewpoint prediction models face limitations due to uncertainties linked with head movement, making it difficult to handle sudden user motions effectively. To overcome these constraints, we propose a lightweight multimodal spatial-temporal transformer architecture, which generates multiple viewpoint trajectories and their corresponding probabilities while leveraging historical trajectory information. Consequently, we introduce a multi-agent reinforcement learning (MARL)-based ABR algorithm that capitalizes on multiple viewport prediction for VR video streaming at the application layer. Our algorithm strives to optimize various QoE objectives under diverse network conditions. To address the ABR problem, we formulate it as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) problem. To tackle this effectively, we develop a MAPPO (Multi-Agent Proximal Policy Optimization) algorithm within a centralized training and decentralized execution (CTDE) framework. Meanwhile, we also improve QoE at the network layer by utilizing network resources in different network nodes during VR video streaming. We present an innovative system called tile-weighted rate-distortion (TWRD) packet scheduling optimization, which takes advantage of viewpoint prediction. The system dynamically assigns weights to tiles and their corresponding packets using the probability of viewpoint prediction. Due to limited bandwidth, the problem of packet scheduling arises, requiring the determination of which packets should be dropped. To address this challenge, we formulate the problem as an optimization task, taking into account error propagation in the video. Our system leverages the weighted rate-distortion information of packets and applies dynamic programming techniques to design an optimal packet scheduling scheme. By selectively dropping packets at network nodes, our proposed system effectively reduces network congestion and enhances the overall performance of VR video streaming systems operating within bandwidth limitations.