Enhanced Equivalence Projective Simulation: A Framework for Modeling Formation of Stimulus Equivalence Classes

Area of Science:

• Computational neuroscience
• Cognitive modeling
• Artificial intelligence

Background:

• Stimulus equivalence class formation is crucial for complex cognition.
• Previous models like Equivalence Projective Simulation (EPS) use episodic memory but have limitations in modeling derived relations.
• Projective Simulation (PS) agents utilize episodic memory akin to cognitive maps for decision-making.

Purpose of the Study:

• To investigate the formation of derived relations in a modified Equivalence Projective Simulation (EPS) model.
• To introduce and evaluate a novel Network Enhancement (NE) process for improved agent decision-making in matching-to-sample tasks.
• To enhance the intrinsic modeling of derived relations and nodal effects compared to existing EPS models.

Main Methods:

• The study employs a modified Projective Simulation (PS) agent with an episodic memory.
• A Network Enhancement (NE) process, an iterative diffusion, is applied post-training to refine the memory network.
• The NE process aims to denoise the memory network, enhancing indirect relations and enabling intrinsic formation and retrieval of derived relations.

Main Results:

• The Network Enhancement (NE) process facilitates the intrinsic formation and retrieval of derived relations within the agent's episodic memory.
• The proposed model, incorporating NE, demonstrates improved intrinsic modeling of derived relations and nodal effects compared to standard EPS.
• Decision-making in the test phase is characterized as a retrieval and update process from the cached memory network, not an ad hoc computation.

Conclusions:

• Network Enhancement (NE) offers a more intrinsic and effective method for modeling derived relations in stimulus equivalence tasks.
• The NE process, interpretable as post-training replay, aligns with findings in behavioral and neuroscience research.
• The developed model provides a robust framework for simulating matching-to-sample behaviors and understanding the role of memory network dynamics.