Reinforcement Learning Critic Networks for Stable Q-Value Estimation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing reinforcement learning algorithms face challenges in efficiently estimating Q-values, particularly in continuous action spaces, leading to instability and overestimation issues.

Innovation Solution

The proposed method, referred to as 'Std Q-target', incorporates a State Action Learned Embedding (SALE) technique and subtracts the standard deviation of Q-target network values from the Q-target network, providing a more conservative and efficient Q-value estimation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If standard reinforcement learning algorithms are used to estimate Q-values in continuous action spaces, then the agent can learn policies, but the Q-value estimation becomes unstable and prone to overestimation

Engineering Contradiction:
ImproveQ-value estimation stabilityVSAvoidQ-value estimation accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent introduces an embedding encoder as an intermediary component that transforms state-action pairs into latent embeddings. These embeddings serve as a mediator between the raw state-action inputs and the Q-value estimation process, enabling the network to learn meaningful representations that improve both stability and accuracy of Q-value estimation.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent modifies the Q-value estimation by changing the parameter representation through learned embeddings. Instead of directly estimating Q-values from raw state-action pairs, the system transforms inputs into latent embedding spaces where the estimation becomes more stable and accurate, effectively changing the parameter space for better performance.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If computationally heavy mechanisms are added to improve Q-value estimation, then estimation accuracy improves, but training time and computational overhead increase

Engineering Contradiction:
ImproveQ-value estimation accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-processing state-action pairs through the embedding encoder to create latent embeddings before Q-value estimation. This preliminary transformation of inputs into meaningful representations accelerates the subsequent estimation process while improving accuracy, as the network operates on pre-processed embeddings rather than raw inputs.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent replaces the traditional mechanical Q-value estimation mechanism with a learned embedding-based approach. Instead of relying on complex computational mechanisms for direct estimation, the system substitutes this with a neural network-based embedding layer that learns efficient representations, reducing computational overhead while improving accuracy.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4535096A1Device and method to improve reinforcement learning
Publication Date: 2025.04.09 ROBERT BOSCH GMBH
  • EP4535096A1 patent drawingFigure 1
  • EP4535096A1 patent drawingFigure 2
  • EP4535096A1 patent drawingFigure 3

AI summary

A computer-implemented method of learning a policy for controlling a computercontrolled system, in particular a robot, by reinforcement learning, comprising the steps of: Observing a current state (s) of an environment of the computercontrolled system. Interacting (S12) with the environment by carrying out the steps of: Determining an action (a) by the policy, which is an actor network, depending on the current state (s), Executing the action (a) by the computercontrolled system, receiving a reward (r) and the next state (s'), and storing the interaction with the environment as an experience in a replay buffer. Sampling (13) a batch of experiences from the replay buffer. Computing (S15) Q values by at least two critic networks for the experiences in said batch. Determining a standard derivation of the Q values for the experiences in said batch for each of the critic networks. Updating (S17) critic parameters depending on the Bellman equation and adding the standard derivations as regularization.