Model AI Assignments 2014 Laura E. Brown Michigan Technological University [email protected]

AAAI 2014

TL;DR

This paper presents a collection of five AI assignments from the 2014 Model AI Assignments session, designed to be easily adoptable, engaging, and flexible for instructors. These assignments aim to provide students with hands-on experience in core AI concepts like Monte Carlo methods, game AI, logical inference, and text classification, fostering deeper understanding and practical skills.

What Problem Does This Paper Address?

The authors address the challenge of effectively teaching Artificial Intelligence (AI) concepts to students. While theoretical knowledge is crucial, practical application through well-designed assignments is essential for solidifying understanding and developing problem-solving skills. The paper highlights the need for assignments that are not only educational but also engaging, flexible for different teaching styles, and easily integrated into existing curricula. The goal is to disseminate high-quality, readily usable assignment designs that can enhance AI education across institutions.

Methods

The paper showcases five distinct AI assignments, each focusing on different areas within AI:

  • An Introduction to Monte Carlo Techniques in AI — Part I: This assignment introduces students to Monte Carlo (MC) simulation and MC reinforcement learning. Students are expected to gain practical experience with these techniques, understand their strengths and weaknesses, and learn to identify appropriate applications. The assignment likely involves implementing MC simulations and potentially reinforcement learning algorithms.
  • Multi-Player Games: Introducing Assignments with Open-Ended Strategies in CS2: This project involves students designing and implementing their own player strategies for commercial board games. The assignment leverages a server-based multi-player game engine and browser-based display, allowing students to focus on strategy development. It emphasizes basic data structures, algorithms, and search techniques, with the added benefit of allowing for continuous strategy improvement, which is motivating for students.
  • Strimko by Resolution: This is a programming-intensive assignment where students build a complete solver for the Strimko puzzle using propositional logic, logical inference via resolution, and backtracking search. The aim is to provide a hands-on, "under the hood" experience with logical reasoning, avoiding black-box SAT solvers. Extra credit opportunities encourage exploration of first-order logic and answer set programming.
  • Party Affiliation Classification from State of the Union Addresses: This assignment focuses on text classification using Naïve Bayes. Students extend their understanding of Naïve Bayes (multinomial and Bernoulli models) by implementing a method to train the classifier and apply it to predict the party affiliation of US Presidents based on their State of the Union Addresses. Extensions can involve text processing techniques and other classification algorithms.
  • Comparing Brute-Force Searching versus the MRV Heuristic in Sudoku Puzzles: This assignment aims to help students understand the benefits of informed search strategies over brute-force search. It uses the number of attempted variable assignments as a criterion to compare brute-force search with an informed search employing the Minimum Remaining Values (MRV) heuristic in Sudoku puzzles. This method provides a measure of efficiency independent of elapsed time, which can be masked by fast hardware or small problem sizes.

Datasets

The paper does not explicitly list or introduce new datasets. However, the assignments imply the use of:

  • Strimko Puzzles: For the "Strimko by Resolution" assignment.
  • State of the Union Addresses (SOTU): For the "Party Affiliation Classification" assignment, used for training and evaluating Naïve Bayes models.
  • Commercial Board Games: Such as Quoridor, San Francisco Cable Cars, Gobblet, and The aMAZEing Labyrinth, for the "Multi-Player Games" assignment.
  • Sudoku Puzzles: For the "Comparing Brute-Force Searching versus the MRV Heuristic" assignment.

Key Insights and Ideas

  • Hands-on experience is crucial for AI learning: The assignments emphasize practical implementation to solidify theoretical concepts.
  • Engaging assignments foster deeper learning: Using puzzles, games, and real-world data makes learning more enjoyable and memorable.
  • Flexibility in assignments benefits instructors: The presented assignments are designed to be adaptable to different course structures and instructor preferences.
  • "Under the hood" understanding is valuable: Avoiding black-box tools in assignments like "Strimko by Resolution" allows students to grasp the fundamental mechanisms of AI techniques.
  • Measuring efficiency beyond elapsed time is important: The Sudoku assignment highlights how to demonstrate the true benefits of informed search algorithms, which can be obscured by hardware speed.
  • Open-ended projects encourage creativity and continuous improvement: The multi-player game assignment allows students to iterate on their strategies, promoting a sense of ownership and mastery.

Key Results

The paper focuses on the pedagogical value and design of the assignments rather than presenting quantitative results from student performance. However, it notes:

  • Positive student reactions: The "Party Affiliation Classification" assignment received positive feedback, with students enjoying applying a classification method to real data.
  • Demonstrations of games: The "Multi-Player Games" presentation will include demonstrations of student-created game strategies.
  • "Look under the hood" of logical reasoning: The "Strimko by Resolution" assignment aims to provide students with a deep understanding of logical inference and backtracking search.
  • Improved understanding of search efficiency: The Sudoku assignment provides a method to better understand the benefits of informed search strategies.

Limitations and Future Work

The paper itself is a collection of assignment abstracts and does not present a single, unified research study with limitations in the traditional sense. However, the authors implicitly acknowledge the following:

  • Need for readily adoptable resources: The existence of the Model AI Assignments session and the provided URL (http://modelai.gettysburg.edu) suggests a continuous effort to gather and disseminate such resources, implying that the development and sharing of high-quality assignments is an ongoing process.
  • Potential for further extensions: Several assignments mention extensions (e.g., "Strimko by Resolution" to first-order logic, "Party Affiliation Classification" to other classifiers), indicating that these assignments can be adapted and expanded for more advanced students or courses.
  • Focus on introductory and intermediate AI: The assignments appear to be tailored for undergraduate AI courses, with some extensions potentially reaching graduate-level concepts. Future work could involve developing assignments for more specialized AI subfields or advanced topics.

Summary updated

BibTeX
@inproceedings{2014model,
  title = {{Model AI Assignments 2014 Laura E. Brown Michigan Technological University [email protected]}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
}
Figure