Predicting Network Events via DFA Rule Sets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models, such as LSTM, are impractical for deployment in hardware-limited network components due to computational resource constraints and unsupported operations, making it difficult to predict transient network events like microbursts and link congestion in real-time.
Innovation Solution
Converting a prediction model into a deterministic finite automaton (DFA) and then into a rule set, such as P4 rules, for deployment on network components, enabling prediction at line rate and microsecond scale without requiring extensive computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models like LSTM are deployed in network components, then prediction accuracy is improved, but hardware resource requirements and computational complexity increase beyond what network components can support
Solution Approach 1:
The patent creates a simplified copy of the LSTM model's predictive behavior using a deterministic finite automaton (DFA). Instead of deploying the complex LSTM model directly in network hardware, the system trains an LSTM model offline, extracts its decision logic, and represents it as a DFA with states, transitions, and outputs. This copy can then be deployed in resource-constrained network components while maintaining prediction accuracy for events like microbursts and congestion.
Solution Approach 2:
The patent extracts the essential predictive logic from the complex LSTM model and separates it from the heavy computational requirements. By taking out only the decision-making structure (the DFA representation) and leaving behind the computationally intensive training and inference processes, the system enables deployment in network components with limited hardware resources.
2Productivity
If complex prediction models are used, then prediction capability is improved, but ease of deployment and operation deteriorates due to resource constraints
Solution Approach 1:
The patent creates a deployable copy of the prediction model in the form of a DFA that can be easily installed in network components. The DFA representation with its finite states and transitions is much simpler to deploy than the original LSTM model, while preserving the prediction capability for network events.
Solution Approach 2:
The patent transforms the model from a continuous, high-dimensional neural network into a discrete, finite-state system. By changing the parameter representation from continuous weights and biases to discrete states and transition rules, the system makes the model suitable for deployment in resource-constrained environments while maintaining predictive performance.
3Device complexity
If deterministic finite automaton is used to represent prediction models, then device complexity is reduced, but manufacturing precision and model fidelity may worsen
Solution Approach 1:
The patent creates a faithful copy of the LSTM model's decision logic in DFA form through systematic extraction of states, transitions, and outputs. This copying process preserves the predictive behavior while simplifying the representation, achieving both model fidelity and reduced complexity.
Solution Approach 2:
The patent uses feedback from the trained LSTM model to guide the DFA construction process. By using the LSTM's predictions and decision boundaries as feedback during state extraction and transition definition, the system ensures that the simplified DFA maintains high fidelity to the original model's predictive capabilities.
Data Source
AI summary
A processing system including at least one processor may obtain a time series of measurement values from a communication network and train a prediction model in accordance with the time series of measurement values to predict future instances of an event of interest, where the time series of measurement values is labeled with one or more indicators of instances of the event of interest. The processing system may then generate a deterministic finite automaton based upon the prediction model, convert the deterministic finite automaton into a rule set, and deploy the rule set to at least one network component of the communication network.