Abstract
In this paper we propose the CTS (Concious Tutoring System) technology, a biologically plausible cognitive agent based on human brain functions. This agent is capable of learning and remembering events and any related information such as corresponding procedures, stimuli and their emotional valences. Our proposed episodic memory and episodic learning mechanism are closer to the current multiple-trace theory in neuroscience, because they are inspired by it contrary to other mechanisms that are incorporated in cognitive agents. This is because in our model emotions play a role in the encoding and remembering of events. This allows the agent to improve its behavior by remembering previously selected behaviors which are influenced by its emotional mechanism. Moreover, the architecture incorporates a realistic memory consolidation process based on a data mining algorithm.
Introduction
In Artificial Intelligence, memory is frequently treated as a static database. However, human cognition relies on episodic memory—the ability to recall the specific context of an event alongside its emotional significance. Existing cognitive architectures, such as ACT-R or LIDA, often struggle with this integration; ACT-R lacks an emotional component for temporal recall, while LIDA’s strict separation of episodic and semantic memory prevents mutual reinforcement. The Conscious Tutoring System (CTS) addresses these gaps by utilizing emotional valences (e.g., satisfaction or fear) to prioritize behaviors and consolidate memories.
The CTS Architecture and Methodology
The CTS architecture is heavily inspired by neurobiology, focusing on the interplay between perception, emotion, and memory consolidation.
Key modules include:
- The Pseudo-Amygdala: This module assigns emotional tags to "coalitions" of information processed by the agent, infusing data with activation levels based on emotional value.
- The Pseudo-Hippocampus: Following the "multiple-trace theory," this module manages episodic memory and facilitates consolidation.
- Sequential Pattern Mining: To manage the combinatorial explosion of raw data, the system uses a modified Generalized Sequential Pattern (GSP) algorithm to extract "frequent closed sequences." This reduces processing time from over an hour to less than a second by focusing only on significant, non-redundant patterns.
- Behavior Selection: The agent selects behaviors based on "sequence strength," a product of a pattern's frequency and the sum of its associated emotional valences.
Experimental Setup
The system was evaluated using the Canadarm2 Simulator, a virtual environment where users learn to manipulate the robotic arm of the International Space Station. The primary dataset consisted of 160 recorded execution sequences of the CTS interacting with students, with each sequence containing approximately 26 cognitive cycles (broadcasts).
Key Results and Performance
Emotions serve as a computational heuristic in CTS, allowing the agent to prioritize behaviors that lead to "self-satisfaction." For instance, the agent learned that providing a subtle hint was more effective than a direct answer, recording a high positive emotional valence (0.8) when the user succeeded following a hint.
Computational efficiency is a hallmark of the system. The episodic learning algorithm runs in under 5 milliseconds. As shown in the performance data, the time required to mine new patterns increases linearly with the number of recorded sequences, staying under 6 seconds even at 160 sequences, which confirms its viability for real-time tutoring applications.
Conclusion and Future Work
The CTS demonstrates that integrating emotional mechanisms into episodic learning allows cognitive agents to prioritize effective behaviors and consolidate memories efficiently. While the current system requires a full re-computation of patterns for new data (the "cold start" problem), future research will focus on incremental mining techniques and empirical studies to measure the direct impact on human student learning outcomes.
Summary updated