Control Agent Training Using Dynamic Models and Action Evaluation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing control agent training methods for technical systems require large amounts of data and are susceptible to poor performance in conditions poorly covered by training data, leading to inefficiencies and prediction errors.

Innovation Solution

A method involving a data-based dynamic model and action evaluation process, such as a variational autoencoder, is used to configure a control agent by minimizing reproduction errors, allowing for efficient training with less data and reducing susceptibility to prediction errors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If large amounts of batch training data are used to train a control agent, then the control agent can cover more operating conditions, but the training data requirement increases and training efficiency decreases

Engineering Contradiction:
Improvecoverage of operating conditionsVSAvoidtraining data volume
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent introduces a replay buffer as an intermediary component that stores and manages training data. This replay buffer acts as a mediator between the environment (technical system) and the neural network, enabling efficient reuse of training data through experience replay. The replay buffer stores transitions (state, action, reward, next state) and samples them in batches for training, reducing the need for continuous collection of large volumes of training data while maintaining comprehensive coverage of operating conditions.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If control actions are restricted through constraints to improve control behavior in poorly covered areas, then control performance improves, but training becomes more data-intensive

Engineering Contradiction:
Improvecontrol behavior in poorly covered areasVSAvoidtraining data volume
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent implements feedback mechanisms through the replay buffer that continuously stores and revisits training experiences. The system samples from the replay buffer to retrains the neural network on previously encountered states and actions, providing feedback loops that improve control behavior in poorly covered areas. This feedback mechanism allows the system to learn from past experiences without requiring additional data collection, thereby improving reliability in underrepresented operating conditions without increasing data volume.

Inventive Principle:
Principle #23Feedback

3Reliability

If Q-function training is used to improve control in underrepresented state space areas, then control behavior improves, but the training process becomes more complex and error-prone

Engineering Contradiction:
Improvecontrol behavior in underrepresented areasVSAvoidtraining process complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent employs experience replay that copies and reuses training data from the replay buffer multiple times for training the neural network. Instead of requiring complex Q-function training mechanisms, the system copies past experiences and retrains the policy network on these copied data samples. This copying approach simplifies the training process while effectively improving control behavior in underrepresented areas by providing repeated exposure to diverse operating conditions without increasing structural complexity.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3940596B1Method for configuring a control agent for a technical system and control device
Publication Date: 2025.10.15 SIEMENS AG
  • EP3940596B1 patent drawingFigure 1~3
  • EP3940596B1 patent drawingFigure 4

AI summary

To configure a control agent (POL), predefined training data is imported, specifying state data records (S), action data records (A), and resulting performance values ​​(R) of the technical system (TS). Using this training data, a data-driven dynamic model (NN) is trained to reproduce a resulting performance value (R) based on a state data record (S) and an action data record (A). Furthermore, an action evaluation process (VAE) is trained to reproduce the action data record (A) based on a state data record (S) and an action data record (A) after information reduction, thereby determining a reproduction error (DR, D0, D1). To train the control agent (POL), training data is fed to the trained dynamic model (NN), the trained action evaluation process (VAE), and the control agent (POL).Performance values ​​(R1, R2) output by the trained dynamic model (NN) are fed into a predefined performance function (P). Reproduction errors (D0, D1) output by the trained action evaluation process (AEP) are also fed into the performance function (P) as performance-reducing factors. This trains the control agent (POL) to output an action data set (A) that optimizes the performance function (P) based on a state data set (S).