Reinforcement Learning Reward Shaping With Expert Progress Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing reinforcement learning systems struggle with tasks that provide extremely sparse rewards and often fail to learn from databases of expert trajectories effectively, particularly when the trajectories contain mistakes made by human experts, leading to suboptimal performance.

Innovation Solution

A reinforcement learning system that trains a policy model neural network using an exploration reward term based on a progress model, which generates a bias to select actions that contribute significantly to task completion, even in the presence of sparse rewards, by utilizing a database of expert trajectories to encourage long-term achievements and filter out uninformative data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If reinforcement learning systems use traditional reward functions with sparse rewards, then the system can operate with simple reward structures, but the system fails to learn effectively from expert trajectories and achieves suboptimal performance

Engineering Contradiction:
Improvelearning effectivenessVSAvoidreward function structure
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The reward function is segmented into two distinct components: an intrinsic reward term that encourages exploration based on progress toward goal states, and an extrinsic reward term that provides sparse task completion feedback. This segmentation allows the system to learn effectively from expert trajectories by providing dense guidance through the intrinsic reward while maintaining the simple sparse structure of the extrinsic reward.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The intrinsic reward term acts as an intermediary between the agent and the sparse extrinsic rewards. It mediates the learning process by providing continuous feedback based on progress toward goal states, enabling the agent to learn from expert trajectories without requiring dense extrinsic rewards at every step.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If the system utilizes databases of expert trajectories to guide learning, then the system can learn from human expertise, but the trajectories contain mistakes made by human experts that lead to suboptimal performance

Engineering Contradiction:
Improvetask performance qualityVSAvoidexpert mistake propagation
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

Instead of having the agent directly imitate expert trajectories and propagate their mistakes, the system inverts the approach by training the agent to reach goal states independently using the intrinsic reward mechanism. The expert trajectories serve as a source of goal states rather than as direct behavior templates, thereby inverting the learning paradigm from imitation to goal-directed exploration that avoids mistake propagation.

Inventive Principle:
Principle #13The other way round (Inversion)

Solution Approach 2:

The mistakes in expert trajectories are converted from a harmful factor into a beneficial one by not using them as direct learning targets. The system benefits from the diversity of goal states in the trajectories while avoiding the propagation of errors, effectively turning the presence of flawed expert data into an opportunity for learning robust goal-directed behavior.

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

3Productivity

If the system focuses on long-term achievements through progress-based rewards, then the system can surpass human experts in fewer time steps, but the system requires complex progress modeling to evaluate task completion progress

Engineering Contradiction:
Improvelearning speedVSAvoidprogress model structure
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system changes the parameter of reward density from sparse (traditional) to dense (progress-based) by introducing the intrinsic reward term. This parameter change enables faster learning and surpassing of human expert performance while the complexity of the progress model is managed through efficient computation of progress metrics based on goal state proximity.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentUS20250348748A1System and method for reinforcement learning based on prior trajectories
Publication Date: 2025.11.13 GDM HOLDING LLC
  • US20250348748A1 patent drawing
  • US20250348748A1 patent drawing
  • US20250348748A1 patent drawing

AI summary

A reinforcement learning system is proposed in which a policy model neural network is trained to control an agent to perform a task in successive time steps, by training a control system including the policy model neural network to select a respective action for each time step which gives a high value for a reward function based on the action, and which indicates the contribution of the action to solving the task. The reward function includes a term based on a progress value output by a progress model. The progress model generates the progress value upon receiving a first observation of the state of the environment at a time step before the performance of the action, and a second observation of the state of the environment at a time step following the performance of the action. The progress value is an estimate of the average time which an ensemble of experts who produced the demonstrations would have taken to transform the environment from how it appears in the first observation to how it appears in the second observation.