Proactive Slice Admission Control

• A proactive slice admission control framework is an advanced network management model that predicts future network demand to optimize long-term objectives, including profit and QoS.
• It integrates deep reinforcement learning, multi-objective heuristics, and dynamic programming to balance conflicting metrics like delay, fairness, and resource efficiency in 5G networks.
• Empirical evaluations show gains such as up to 15% profit increases and 30% delay reductions, highlighting its practical benefits in NFV-based substrate environments.

A proactive slice admission control framework is an advanced architectural and algorithmic solution for maximizing long-term network provider objectives—such as profit, QoS guarantee, fairness, and resource efficiency—by making forward-looking slice admission decisions in 5G and beyond network slicing scenarios. These frameworks depart from myopic or purely reactive mechanisms by explicitly embedding predictions, delay-aware incentives, resource reservation, forecasting, or long-horizon optimization into the controller's logic. The goal is to optimally balance conflicting objectives (e.g., profit vs. delay, priority vs. fairness, capacity vs. future demand) under real-world settings that include stochastic arrivals, heterogeneous QoS constraints, resource coupling, and rapidly varying operating conditions (Chakraborty et al., 9 Oct 2025).

1. System Architectures and MDP Formulations

Proactive slice admission control frameworks typically operate atop NFV-based substrate networks, mediating between incoming network slice requests and the multi-dimensional physical resource pool (compute, bandwidth, storage). Core modules generally include:

• Slice Queue Manager (Prioritizer): Aggregates and prioritizes incoming slice requests (e.g., eMBB, URLLC, mMTC) into a queue.
• Admission Controller (Policy Agent): Observes the network state and queue composition, selecting which requests to admit based on optimized policy, often using learned or computed priority vectors.
• Resource Pool and Allocator: Tracks available resources on substrate nodes and links, possibly incorporating prediction for release and arrival events, and attempts embedding/admission in priority order, updating the system state accordingly (Chakraborty et al., 9 Oct 2025, Dai et al., 2022, Hoang et al., 2017).

Formulation as Markov Decision Processes (MDP) or Semi-Markov Decision Processes (SMDP) is standard. The system state encodes substrate resources, current slice occupancy, and queue composition, while actions correspond to admission/rejection or assignment of priority weights. Transition dynamics capture the stochastic evolution due to arrivals, admissions, departures, and resource releases. Many works incorporate multi-queue models and continuous or discrete-time state descriptions (Chakraborty et al., 9 Oct 2025, Han et al., 2019, Tao et al., 2023).

2. Core Algorithmic Components: DRL, Multi-Objective Heuristics, and Stochastic Control

A spectrum of algorithms underlies proactive SAC frameworks.

Deep Reinforcement Learning (DRL):

• Double-DQN architectures with feed-forward neural networks are employed, where inputs are substrate/resource states and queued slice features; outputs are Q-values for discrete admission actions (e.g., vector of priority weights).
• Delay-aware reward functions penalize delay violation for latency-critical slices, combining normalized profit and explicit delay penalties:

$$R_t = \alpha\,\mathrm{Profit}_t - \beta \max(0, \mathrm{Delay}_t - D_\mathrm{max})$$

• Exploration is handled by Boltzmann (softmax) policies for stability and fast convergence, as opposed to less stable $\epsilon$-greedy methods (Chakraborty et al., 9 Oct 2025).
• Digital Twin–assisted DRL initializes and accelerates the learning process by bootstrapping with a deterministic policy model (Tao et al., 2023).

Multi-Objective Heuristics:

• Resource-efficiency–based priority adjustment calculates per-slice-type marginal CSAR gain per unit resource, ensuring cumulative service acceptance ratios (CSARs) satisfy priority and fairness constraints.
• Target CSAR tracking guides fair resource allocation, tuning the fairness–priority trade-off via scalar thresholds.
• Two-phase approaches first adjust priority monotonicity, then enforce fairness by allocating to underserved types, subject to resource constraints (Dai et al., 2022).

Stochastic Dynamic Programming and Prediction:

• Value iteration and Bellman optimality yield policies that trade immediate reward against expected future returns, protecting future headroom for high-value or priority slices (Hoang et al., 2017, Han et al., 2018).
• Proactive frameworks use traffic forecasting or Markov transition models to anticipate overload or QoS bottlenecks, either through model-driven or learned predictors (Han et al., 2018, Jacoby et al., 9 Jan 2026).
• Admission controllers may use multi-dimensional knapsack–style online algorithms that dynamically adjust acceptance thresholds as a function of evolving resource scarcity or predicted demand, with $O(m)$ per-request complexity (Ajayi et al., 8 Aug 2025).

3. Delay and QoS-Aware Reward Design

A distinguishing feature of modern proactive admission control is the explicit incorporation of delay-awareness and other QoS penalties into the reward or objective function:

• Penalties for exceeding latency bounds, especially for services such as URLLC, are subtracted from the instantaneous profit,
• Rejection penalties discourage frivolous declines by imposing a negative reward of order comparable to a typical slice's profit (Chakraborty et al., 9 Oct 2025).
• Normalization over theoretical profit maxima bounds reward values for stable training.

This design forces the agent to admit slices in a way that balances immediate revenue and the long-term degradation from delay-sensitive SLA violations, leading to policies that prioritize low-latency slices while still maintaining high resource utilization and profit.

4. Evaluation, Empirical Performance, and Practical Guidance

Empirical evaluations, commonly using large-scale, synthetic or realistic topologies, quantify the following metrics:

• Normalized NSP profit
• Average per-type (e.g., URLLC) slice delay
• Acceptance rate of admitted slices
• Resource utilization by type (CPU, bandwidth)

In direct comparisons, proactive frameworks (e.g., DePSAC) achieve:

• Profit increases up to 15%,
• URLLC delay reductions up to 30%,
• Acceptance rate improvements of 10 percentage points,
• Bandwidth consumption reductions while maintaining CPU usage (Chakraborty et al., 9 Oct 2025).

Convergence is accelerated and oscillatory training behavior is mitigated by softmax-based exploration and carefully tuned trade-off coefficients ($\alpha$, $\beta$, $\gamma$). A practical tuning regime is to match the delay penalty coefficient $\beta$ to the revenue per unit delay violation, and to set the rejection penalty to approximately a single-slice profit (Chakraborty et al., 9 Oct 2025).

Scalability concerns arise with fine-grained action or state representations; policy gradient methods with continuous actions or action quantization may be required for very large networks.

5. Proactive Resource Reservation, Fairness, and Extensions

Frameworks integrate resource reservation logic to guarantee future capacity for high-priority or delay-critical requests:

• Action spaces and prioritization queues may encode "headroom" by temporarily deferring, or even intentionally rejecting, low-priority slices if future predicted demand or MDP planning indicates an anticipated bottleneck (Hoang et al., 2017, Chakraborty et al., 9 Oct 2025, Han et al., 2019).
• Fairness is handled by tracking and targeting monotonicity in acceptance or service ratios among slice types, a non-trivial issue given underlying resource contention and heterogeneity. Dynamic tuning of fairness versus priority parameters enables flexible service-level differentiation (Dai et al., 2022).
• Resource allocation may be coupled with online learning of arrival and service patterns to robustify control to nonstationary conditions.

6. Limitations and Open Challenges

• Most proactive SAC frameworks assume either stationary traffic patterns or rely on retraining or online adaptation to cope with nonstationarity; their efficacy under highly bursty or adversarial load remains an active research area.
• The size of the state and action space scales rapidly with network complexity and prioritization granularity, motivating continued research in scalable function-approximation or hierarchical control.
• Interactions between proactive admission, resource mapping, and ongoing congestion control require integrated frameworks that reason over multiple time scales.

7. Representative Framework Comparison Table

Framework	Main Technique	Delay/QoS Awareness	Prioritization	Fairness Support	Empirical Gains
DePSAC (Chakraborty et al., 9 Oct 2025)	DQN+Boltzmann DRL	Explicit delay penalty	Yes	No	+15% profit, –30% delay
PSACCF (Dai et al., 2022)	Multi-objective heuristic	No	Yes	Yes	+33.6% fairness, +9% util
Value Iteration (Hoang et al., 2017)	MDP value iteration	Indirect	Yes	No	2–3× reward over greedy
OSAC (Ajayi et al., 8 Aug 2025)	Reservation-based online knapsack	No	By price	No	+12.9% revenue

This table summarizes core properties and observed outcomes in representative frameworks.

---

Proactive slice admission control is thus an essential paradigm in 5G and beyond, integrating delay- and priority-sensitive objectives, predictive resource management, and advanced optimization or learning algorithms, with demonstrated performance and QoS gains across a wide spectrum of slicing scenarios (Chakraborty et al., 9 Oct 2025Dai et al., 2022Hoang et al., 2017Ajayi et al., 8 Aug 2025).