AI Papers Reader

Personalized digests of latest AI research

View on GitHub

AITEE: AI Tutor Tackles Electrical Engineering Education with Agentic Approach

Researchers have unveiled AITEE, a novel intelligent tutoring system designed to revolutionize electrical engineering education. AITEE aims to address the “teacher-bandwidth problem” by providing scalable, personalized, and effective learning environments, especially in the challenging domain of electrical circuit analysis.

The core challenge lies in the complexity of understanding and solving electrical circuit problems. First-semester electrical engineering students must grasp mathematical foundations like linear algebra, and also electrical engineering principles such as Kirchhoff’s laws. AITEE tackles this by combining several technologies including image processing of circuit diagrams (both digital and hand-drawn), graph neural networks (GNN) for representing circuit structure, and large language models (LLMs) augmented with retrieval-augmented generation (RAG) for contextual knowledge.

Here’s how AITEE works:

  1. Circuit Recognition: AITEE can process images of electrical circuits, whether they are computer-generated or hand-drawn. This involves identifying components (resistors, voltage sources, etc.) and their connections. For example, even a hand-drawn circuit with a star configuration can be recognized.

  2. Graph Representation: The circuit is transformed into a graph structure, where components and connection points are nodes, and connections are edges. This graph encapsulates the circuit’s topology. A graph neural network (GNN) then generates a vector representation of this graph. This representation is called a “circuit embedding”.

  3. Contextual Knowledge Retrieval: When a student asks a question (e.g., “How do I calculate the current I3?”), AITEE uses the circuit embedding of the student’s circuit to retrieve relevant context from lecture materials. This is done using a graph-based similarity measure (specifically, the cosine similarity between circuit embeddings). For instance, if the student’s circuit is a complex mixed circuit, AITEE will pull information from lecture sections on mixed circuits and potentially simplify its analysis into a series circuit.

  4. LLM-Powered Tutoring: A large language model (LLM), augmented with the retrieved context, engages the student in a Socratic dialogue. This means the LLM asks guiding questions to help the student arrive at the solution independently, rather than simply giving the answer. Imagine the LLM asking, “What are Kirchhoff’s laws, and how might they apply here?”

  5. SPICE Simulation for Accuracy: To ensure accurate answers, AITEE utilizes SPICE simulation to independently verify results, preventing “AI hallucinations” (factually incorrect responses) that can be detrimental to learning. So, before telling a student that the current is 2 amps, it confirms that through SPICE.

The researchers evaluated AITEE using different LLMs and various prompting strategies. They found that AITEE significantly outperforms baseline approaches in domain-specific knowledge application, with even medium-sized LLM models showing acceptable performance. The effectiveness of the graph representation for circuit topology was clearly demonstrated, with similar circuits having embeddings with a cosine similarity approaching 1.

However, the authors note that challenges remain, particularly improving robustness against inaccurate user input and enhancing dialogue to better foster learner autonomy. They’re also looking to reduce errors performing arithmetic operations by integrating a simulation execution capability using PySpice. Future work involves testing AITEE in real-world educational settings to assess its impact on student learning outcomes.