Reinforcement Learning Agent Exploration via Observation Likelihood Divergence

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing reinforcement learning systems are inefficient in exploring environments due to sensitivity to noise and require excessive computational resources, limiting their ability to discover new information and perform tasks effectively.

Innovation Solution

A reinforcement learning system that trains a neural network using a reward function based on the difference between two statistical models of environmental interactions, encouraging actions that maximize informative observations, and adapts the network parameters to optimize exploration and task performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If existing reinforcement learning systems are used to explore environments, then the agent can interact with the environment and collect data, but the system exhibits high sensitivity to noise and requires excessive computational resources

Engineering Contradiction:
Improvenoise sensitivityVSAvoidcomputational resource usage
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent introduces statistical models as intermediary components that mediate between the agent's observations and the reinforcement learning decision-making process. These statistical models pre-process and structure the environmental observations, filtering out noise before it reaches the main RL algorithm. This intermediary layer reduces the computational burden on the primary RL system while improving robustness to noisy observations, directly addressing the contradiction between noise sensitivity and computational resource usage.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Loss of information

If existing reinforcement learning systems explore environments extensively, then the agent can discover new information, but the process requires excessive computational resources and time

Engineering Contradiction:
Improveinformation discoveryVSAvoidexploration efficiency
Core Design Contradiction:
Loss of informationVSProductivity

Solution Approach 1:

The patent implements feedback mechanisms where the statistical models continuously learn from and adapt to the agent's interactions with the environment. The models use the collected observations to update their understanding of environmental dynamics, which then feeds back into guiding future exploration actions. This feedback loop enables more efficient information discovery by leveraging learned patterns to prioritize promising exploration directions, reducing the computational resources and time needed compared to pure random or systematic exploration methods.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The statistical models perform preliminary analysis and structuring of environmental data before the main reinforcement learning algorithm processes it. By pre-computing statistical representations and patterns from observations, the system prepares processed information in advance that guides more efficient exploration decisions. This preliminary action reduces the computational workload during actual exploration and accelerates the discovery of new information.

Inventive Principle:
Principle #10Preliminary action

3Loss of information

If the neural network is trained to maximize informative observations, then the agent can learn about the environment more effectively, but the training process requires complex reward function calculations

Engineering Contradiction:
Improveinformation gainVSAvoidreward function complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The statistical models serve as intermediaries that simplify the calculation of informative observations for the reward function. Instead of directly computing complex information-theoretic measures from raw observations, the statistical models provide pre-processed representations that make it easier to quantify information gain. This intermediary representation layer reduces the computational complexity of the reward function while maintaining its ability to effectively guide the agent toward informative observations.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3915053B1Controlling an agent to explore an environment using observation likelihoods
Publication Date: 2024.07.31 DEEPMIND TECH LTD
  • EP3915053B1 patent drawingFigure 1
  • EP3915053B1 patent drawingFigure 2
  • EP3915053B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for controlling an agent. One of the methods includes, while training a neural network used to control the agent, generating a reward value for the training as a measure of the divergence between the likelihood of the further observation under first and second statistical models of the environment, the first statistical model and second model being based on respective first and second histories of past observations and actions, the most recent observation in the first history being more recent than the most recent observation in the second history.