Action Selection Rule Generation for Technical System Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex technical systems like gas turbines and wind turbines often rely on computer-aided control methods using neural networks for action selection rules, which result in high complexity, making them difficult for human experts to interpret and understand, leading to suboptimal control due to lack of comprehensibility.
Innovation Solution
A method is developed to generate and select action selection rules with lower complexity by assigning a complexity measure and evaluating them based on distance, reward, and quality measures, using techniques like genetic programming and particle swarm optimization, ensuring the rules are understandable and effective for control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If complex regression methods such as neural networks are used to determine action selection rules, then the optimality of control is improved, but the complexity of the action selection rule increases making it incomprehensible to human experts
Solution Approach 1:
The patent segments the action selection process by separating the complex neural network evaluation (performed offline to generate training data) from the actual control decision-making (performed online using simplified rules). This allows the complex computations to be done once during training while the operational system uses comprehensible, simple rules for real-time control.
Solution Approach 2:
The patent creates a simplified copy or approximation of the optimal control policy derived from complex neural networks. Instead of using the full complex model for operational decisions, it generates training data from the complex model and uses this to train simpler action selection rules that replicate the essential control behavior while remaining comprehensible.
2Device complexity
If simpler regression methods or low-dimensional state representations are used, then the complexity of the action selection rule is reduced for better comprehensibility, but the optimality of control deteriorates
Solution Approach 1:
The patent performs preliminary actions by using complex neural networks offline to generate high-quality training data before deploying the simplified control system. This preliminary computation phase captures the optimal control behavior, which is then encoded into simpler rules that can be executed efficiently during operation without sacrificing control quality.
Solution Approach 2:
The patent introduces training data as an intermediary between the complex neural network and the simple action selection rules. The complex model generates training data that serves as a bridge, allowing the simple rules to learn optimal control strategies without directly containing the complexity of the neural network architecture.
3Productivity
If complex action selection rules are generated using neural networks, then the control performance is optimized, but the ease of operation decreases due to lack of comprehensibility
Solution Approach 1:
The patent segments the control system into two distinct parts: an offline training phase using complex neural networks to generate optimal control data, and an online execution phase using simple, comprehensible rules. This segmentation allows high control performance to be achieved during training while maintaining ease of operation during actual control execution.
Solution Approach 2:
The patent substitutes the mechanical/neural network-based complex decision-making system with a simpler rule-based system for operational use. The complex neural network is replaced by simplified action selection rules that are easier for human experts to understand and operate, while maintaining control performance through careful training data generation.
Data Source
Figure 1
AI summary
The invention concerns a method for the computerized control and/or regulation of a technical system. Within the context of the method according to the invention, an action-selection rule (PO') is determined which has a low level of complexity and yet is well suited to the regulating and/or control of the technical system, there being used for determination of the action-selection rule (PO') an evaluation measure (EM) which is determined on the basis of a distance measure and/or a reward measure and/or an action-selection rule evaluation method. The action-selection rule is then used to control and/or regulate the technical system. The method according to the invention has the advantage of the action-selection rule being comprehensible to a human expert. Preferably, the method according to the invention is used for regulating and/or controlling a gas turbine and/or a wind turbine.