Dynamic network architectures that utilize communication, computing, and storage resources at the wireless edge are key to delivering emerging services in next-generation networks (e.g., AR/VR, 3D video, intelligent cars, etc). Network slicing can be significantly enhanced by including dynamically available resources throughout the fog/edge/cloud continuum and using mmWave/THz bands. However, network slicing of dynamic multi-tier computing networks remains under-explored. In this paper, we present a self-learning end-to-end network slicing mechanism (SELF-E2E-NS) that facilitates collaboration between the Infrastructure Provider (InP) and tenants to slice their subscribers’ resources (i.e., radio, computing, and storage) as fog resources. To adapt to the uncertain availability of resources at the edge and minimize the risk of non-satisfying service level agreements (SLAs), our slicing mechanism has two operational modes. Operational mode 1 is for joint network slicing (JNS) in which the InP infrastructure is augmented with fog resources and jointly sliced to meet high throughput and delay tolerant requirements. Operational mode 2 is for independent network slicing (INS) in which the InP infrastructure and fog resources are sliced separately to achieve high throughput, low-latency, and high-reliability requirements. Our schemes leverage mmWave/THz, fog/edge/cloud computing, and caching to achieve new service requirements. We design a DQ-E2E-JNS algorithm that uses Deep Dueling network and a MAAC-E2E-INS algorithm based on multi-agent actor-critic, which incorporate service-aware pricing feedback and fog trading matching, respectively. These algorithms find the optimal slice request admission and collaboration policy that maximizes the long-term revenue of the InP and tenants for each mode. The simulation results show that our novel slicing mechanism can serve up to 4 times more requests and effectively exploits different spectrum bands and fog resources to improve re...