Stochastic automata-based estimators for adaptively compressing files with nonstationary distributions

Area of Science:

• Machine Learning
• Information Theory
• Data Compression

Background:

• Traditional data compression methods struggle with nonstationary data distributions.
• Weak estimators have shown promise in adaptive signal processing.
• Existing adaptive coding schemes often rely on maximum likelihood or Bayesian estimation.

Discussion:

• This work introduces learning automata for adaptive data compression in nonstationary environments.
• The proposed adaptive coding scheme uses stochastic learning-based weak estimation to update symbol probabilities without traditional methods.
• The approach is integrated into adaptive Fano coding and an entropy-based scheme similar to arithmetic coding.

Key Insights:

• Learning automata effectively handle nonstationary data distributions in compression.
• The novel adaptive schemes outperform their maximum likelihood-based counterparts by nearly 10%.
• This method avoids complex computations associated with maximum likelihood, Bayesian, or sliding-window estimators.

Outlook:

• Potential for improved compression ratios in real-world, dynamic data.
• Further research into optimizing learning automata parameters for diverse data types.
• Exploration of these techniques in other adaptive signal processing applications.