A Self-Organizing Incremental Spatiotemporal Associative Memory Networks Model for Problems with Hidden State.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Identifying hidden states is crucial for solving problems in partially observable environments.
• Deterministic partially observable Markov decision processes (POMDPs) present unique challenges in state estimation.

Purpose of the Study:

• To propose a method for representing any deterministic POMDP using a minimal, looping hidden state transition model.
• To introduce a novel Spatiotemporal Associative Memory Network (STAMN) for realizing this model.
• To demonstrate the efficacy of STAMN in identifying and recalling hidden states.

Main Methods:

• Development of a heuristic algorithm for constructing minimal, looping hidden state transition models.
• Introduction of the Spatiotemporal Associative Memory Network (STAMN).
• Utilizing neuroactivity decay for short-term memory and connection weights for long-term memory within STAMN.
• Employing presynaptic potentials and synchronized activation for simultaneous state identification and recall.

Main Results:

• Demonstrated that any deterministic POMDP can be represented by a minimal, looping hidden state transition model.
• Empirical illustrations validated the functionality of the proposed STAMN.
• STAMN showed competitive performance compared to existing methods in hidden state identification tasks.

Conclusions:

• The proposed STAMN provides an effective mechanism for modeling and identifying hidden states in POMDPs.
• The neuro-inspired design of STAMN, incorporating memory decay and associative recall, offers a promising approach for complex sequential decision-making problems.
• This work contributes to advancing AI's ability to handle uncertainty and hidden information.