AlphaSeq

AlphaSeq applies deep reinforcement learning to discover sequences with desired properties by modeling sequence construction as an episodic symbol-filling Markov decision process.

Key Features:

• Episodic Symbol-Filling Game: Models sequence discovery as an episodic game where symbols are sequentially filled into vacant positions and completed sequence sets are evaluated against predefined criteria.
• Markov Decision Process (MDP) Framework: Frames the problem as an MDP and employs a deep reinforcement learning framework inspired by AlphaGo to make sequential decisions during sequence construction.
• Progressive Learning: Learns through repeated episodes, iteratively improving its policy for generating sequences that optimize complex objectives.
• Suitability for Complex Objectives: Addresses objectives that are analytically intractable or too complex for traditional mathematical tools by navigating large solution spaces via learning.

Scientific Applications:

• Multi-Carrier Code-Division Multiple Access (CDMA) Systems: Rediscovered a set of ideal complementary codes capable of zero-forcing all potential interferences.
• Pulse Compression Radar Systems: Identified sequences that improved the signal-to-interference ratio for mismatched filter (MMF) estimators, outperforming Legendre sequences.

Methodology:

Models sequence discovery as an episodic symbol-filling game within an MDP and applies AlphaGo-inspired deep reinforcement learning to iteratively evaluate completed sequence sets against predefined criteria.