Detecting Changes and Avoiding Catastrophic Forgetting in Dynamic Partially Observable Environments.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Reinforcement Learning

Background:

• Adapting to environmental changes is crucial for intelligent agents.
• Partially observable environments (POMDPs) pose challenges for state detection.
• Existing predictive learning methods often focus on static environments.

Purpose of the Study:

• To develop a novel algorithm for agents to detect and adapt to changes in partially observable environments.
• To enable continuous learning and prevent catastrophic forgetting when encountering new environments.
• To provide explainable AI by justifying model beliefs based on statistical evidence.

Main Methods:

• Proposed an algorithm using statistical tests to assess model fit to the current environment.
• Exploited underlying probability distributions of predictive models for fast and explainable belief assessment.
• Enabled continuous training of separate models for different environments by labeling incoming data.

Main Results:

• The novel method prevents catastrophic forgetting when new environments or tasks are encountered.
• Demonstrated empirical benefits in a set of POMDP simulations.
• The approach provides a fast and explainable method for assessing model beliefs.

Conclusions:

• The proposed algorithm enhances AI agent adaptability in dynamic, partially observable environments.
• Statistical testing of predictive models offers a robust solution for continuous learning.
• The method supports AI-informed decision-making with verifiable justifications.