Actuator Control Policy Using UCB Advantage Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional model-free reinforcement learning methods require extensive exploration and costly human demonstrations to teach robots new tasks, which is inefficient and resource-intensive, especially when environment models or robot dynamics are unknown.

Innovation Solution

The method employs a series of waypoints defined by states of the actuator or its environment, using multiple independent artificial neural networks to determine optimal actions based on upper confidence bounds, allowing for efficient exploration and learning without a correct environment model, by prioritizing actions that maximize long-term returns and avoiding collisions or unreachable states.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If conventional model-free reinforcement learning is used to learn tasks without prior knowledge, then the agent can learn any task, but millions of examples are needed before the robot discovers the task's goal

Engineering Contradiction:
Improveability to learn tasks without prior knowledgeVSAvoidtime required to discover task goal
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The system performs preliminary action by defining a sequence of waypoints that guide the robot toward the goal before actual task execution. These waypoints are calculated based on the current state and desired goal state, allowing the robot to efficiently navigate toward the goal without requiring millions of trial-and-error examples. The waypoints serve as pre-planned intermediate targets that accelerate learning.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system introduces waypoints as an intermediary between the current state and the final goal state. Instead of directly attempting to reach the goal through random exploration, the robot uses waypoints as intermediate targets that guide its exploration. This intermediary structure significantly reduces the number of examples needed by providing a structured path toward the goal.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If human experts demonstrate the task to reduce exploration cost, then the agent learns faster, but time, costs, and human oversight resources are required

Engineering Contradiction:
Improvelearning speedVSAvoidhuman oversight resources
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system implements self-service by automatically generating waypoints and exploration strategies without human intervention. The waypoint generation algorithm autonomously computes a sequence of intermediate targets based on the current state and goal state, allowing the robot to learn tasks independently without requiring human experts to demonstrate each task. This eliminates the need for human oversight resources while maintaining fast learning speed.

Inventive Principle:
Principle #25Self-service

3Ease of operation

If the exploration assumes the environment is free of obstacles, then the exploration process is simplified, but collisions may occur that abort movement to waypoints

Engineering Contradiction:
Improveexploration process simplicityVSAvoidmovement reliability to waypoint
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The system applies partial action by attempting to move to waypoints under simplified assumptions (environment free of obstacles), but only partially executing the movement if a collision is detected. When a collision occurs, the system aborts the current waypoint movement and selects the next waypoint in the sequence, effectively trying multiple partial actions until successful navigation is achieved. This approach maintains simplicity while handling unreliable movements.

Inventive Principle:
Principle #16Partial or excessive action

4Productivity

If multiple independent artificial neural networks are used to determine advantages, then the exploration efficiency is improved, but the computational complexity increases

Engineering Contradiction:
Improveexploration efficiencyVSAvoidnumber of neural networks
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system segments the advantage determination function into multiple independent artificial neural networks, each responsible for evaluating different aspects of potential actions. Instead of using a single complex network, the patent divides the evaluation task across multiple simpler networks that independently compute advantages for different actions. This segmentation improves exploration efficiency by enabling parallel evaluation while keeping individual network complexity manageable.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3741518B1Method and device for automatically influencing an actuator
Publication Date: 2022.09.21 ROBERT BOSCH GMBH
  • EP3741518B1 patent drawingFigure 1
  • EP3741518B1 patent drawingFigure 2
  • EP3741518B1 patent drawingFigure 3

AI summary

Device (100) and method for automated control of an actuator (108), in particular a robot, a machine, an at least partially autonomous vehicle, a tool or a part thereof, wherein at least one state of the actuator (108) or an environment of the actuator (108) is provided by an exploration strategy for learning a policy (300), wherein an action for the automated control of the actuator (108) is defined by the policy depending on the state (308), wherein a state value is defined as an expected value for a sum of rewards that are achieved when following the policy starting from the state, and wherein a state action value is defined as an expected value for a sum of rewards that are achieved when any action is first performed in the state and then the policy is executed.wherein an advantage is defined as a difference between the state value and the state-action value, wherein, depending on the action and the state, a plurality of advantages is defined by a plurality of independent artificial neural networks, wherein the policy for the state defines the action that maximizes an empirical mean over a distribution of the plurality of advantages, wherein the exploration strategy specifies the at least one state that locally maximizes an upper confidence bound, wherein the upper confidence bound is defined as a function of an empirical mean and a variance over the distribution of the plurality of advantages.