An adaptive neural network admission controller for dynamic bandwidth allocation.

Area of Science:

• Computer Science
• Telecommunications Engineering
• Network Management

Background:

• Hybrid-switching networks dynamically share output link bandwidth between isochronous (guaranteed) and asynchronous (best-effort) traffic.
• Admission controllers aim to minimize connection request refusal rates and packet loss in finite buffers.
• Accurate modeling of connection dynamics is crucial for effective bandwidth management.

Purpose of the Study:

• To develop and evaluate an admission control strategy for hybrid-switching networks using artificial intelligence.
• To approximate optimal bandwidth allocation strategies, particularly for scenarios with unknown or slowly varying input rates.
• To minimize both connection refusal rates and packet loss probabilities.

Main Methods:

• Utilized backpropagation feedforward neural networks to approximate optimal admission control strategies.
• Modeled the connection dynamics using an embedded Markov chain.
• Considered scenarios with unknown and slowly varying input traffic rates.
• Compared neural network approximation with dynamic programming for optimal solutions.

Main Results:

• Neural networks effectively approximated optimal admission control strategies.
• The proposed method demonstrated proficiency in minimizing refusal rates and packet loss.
• Performance was validated across various traffic conditions, including dynamic input rates.
• Approximation accuracy was comparable to optimal solutions derived from dynamic programming.

Conclusions:

• Backpropagation feedforward neural networks offer a viable and efficient method for admission control in hybrid-switching networks.
• The approach successfully balances the competing objectives of minimizing connection refusals and packet loss.
• This AI-driven strategy is particularly effective in dynamic network environments with unpredictable traffic patterns.