A new method of Bayesian causal inference in non-stationary environments

Area of Science:

• Artificial Intelligence
• Machine Learning
• Statistical Inference

Background:

• Traditional Bayesian inference requires extensive data for accurate causal estimation.
• Dynamic environments necessitate methods like exponential moving average (EMA) to adapt to changing data.
• EMA introduces a trade-off: higher discounting improves responsiveness but reduces accuracy.

Purpose of the Study:

• To propose an Extended Bayesian Inference (EBI) method that integrates human-like causal inference.
• To enhance Bayesian inference with learning and forgetting effects for dynamic environments.
• To evaluate EBI's performance against existing methods in a changing data scenario.

Main Methods:

• Incorporation of human-like causal inference into the Bayesian inference framework.
• Development of Extended Bayesian Inference (EBI) to introduce learning and forgetting effects.
• Comparative evaluation of EBI against EMA and sequential discounting expectation-maximization algorithms using a dynamic Gaussian mixture model.

Main Results:

• Extended Bayesian Inference (EBI) successfully incorporates learning and forgetting effects.
• EBI demonstrates the ability to modify the inherent trade-off between followability and accuracy observed in EMA.
• Performance evaluation shows EBI's effectiveness in a dynamically changing data environment.

Conclusions:

• Extended Bayesian Inference (EBI) offers an improved approach to causal inference in dynamic systems.
• EBI provides a novel way to balance responsiveness and accuracy, outperforming traditional EMA.
• The integration of causal inference enhances Bayesian methods for real-world, evolving data.