Improving the Teaching of Artificial Intelligence Through Project-Based Learning on a Board Game

Wed, 01 Jan 2025 00:00:00 +0000

Research setting

The paper studies how to teach Artificial Intelligence (AI) to master students with a practical project. It applies Project-Based Learning (PBL), a teaching approach in which students learn by working, alone or in groups, on a real problem that ends in a product. Here the problem is a game: each team writes an AI player for Tablut, an ancient Nordic board game, and the players then face each other in a competition. Tablut is played on a 9 x 9 board by two players with different roles, and each AI player must be able to play both roles.

Tablut suits a first AI course for three reasons given in the paper. Both players see the whole board and no move depends on chance, so simple state-space search, the exploration of possible future board positions, can play it. It is asymmetric, because the two sides start with different checkers and goals, as the board at the top of this page shows. It is also little studied and still unsolved, so students cannot copy a known solution.

Board state The player receives the current position.
Decide For example, by searching possible future positions.
Legal move It returns a move that follows the rules.

What an AI player for Tablut does at each turn, following the requirements in Section 4.2 of the paper.

Motivation

University AI lessons are mostly lectures, which leave students few chances to put theory into practice. Competitions can raise interest and motivation. They can also cause stress and performance anxiety, or draw attention away from the regular course. An earlier edition of this experience used the game Nine Men’s Morris, and its students asked for a less-known game and for freedom in the programming language.

Study	Asymmetric game	Any programming language	Courses and iterations	Students analyzed	Remote lectures analyzed
Earlier edition on Nine Men’s Morris (Chesani et al., 2017)	no	no	1 course, 3 iterations	36	not discussed
This paper (Tablut competition) This work	yes	yes	2 courses, 6 iterations	166 answers	yes

How this study extends the earlier edition of the experience, as described in Sections 2.2 and 4.1 of the paper.

Objective. Improve the learning experience of AI students, such as their motivation and interest, through a project run as a game competition, and measure its effects on knowledge, motivation, soft skills, and the competitive aspect.

Approach

The experience runs within the course Fundamentals of Artificial Intelligence, for a Master Degree in Computer Engineering and a Master Degree in AI at the University of Bologna. Participation is voluntary, in teams of at most 4 students, and finishing the activity gives a bonus of about 7% on the final course score.

Launch Introduced about 4 weeks into the course.
Build a player Teams code their AI player in 6 to 8 weeks.
Validation Staff checks that every player works.
Matches Two round-robin groups, then a final.
Results lesson Teams present their work, and winners are announced.

How one semester of the experience unfolds (Section 4.2 of the paper).

In the round-robin groups, each player meets every other player twice, once as attacker and once as defender. Each game runs as three processes: two players and a Java server that acts as game engine and referee. Players run inside virtual machines with the same fixed resources, which makes the comparison fair.

How a game is run (Section 4.3 of the paper). JSON messages let each team use the programming language it prefers.

To measure the effects, students fill in an anonymous, voluntary questionnaire every semester. It has 4 groups of 5 statements: knowledge of AI (A), interest and motivation (B), personal skills (C), and competition (D). During the COVID-19 pandemic, a fifth group (E) asked about remote lectures. Students rate each statement on a Likert scale from 1 (strongly disagree) to 5 (strongly agree). A t-test checks whether the answers are statistically significant, and two education researchers group the answers to an open question into categories.

Results

The experience ran for 6 semesters, from 2019 to 2021. The results are the average Likert score and the standard deviation for each statement, over all questionnaire answers.

244students in 85 teams over 6 semesters
166questionnaire answers, about 70% of participants
4.46agreement that competing promotes motivation and interest (out of 5)

Average agreement (Likert scale) (higher is better)

A1 Consolidates theoretical AI concepts 4.34 ± 0.73
A4 Teaches new programming concepts 3.63 ± 1.08
B1 Course assessment before the experience 3.93 ± 0.87
B2 Course assessment after the experience 4.22 ± 0.77
C5 Teaches cooperation and teamwork 4.30 ± 0.95
D1 Competing improves team collaboration 4.23 ± 0.89
D2 Competing improves collaboration with other teams 2.85 ± 1.19
D3 Competing promotes motivation and interest 4.46 ± 0.83
D4 Participation is linked to winning the first prize 2.45 ± 1.21

Average Likert score from 1 (strongly disagree) to 5 (strongly agree) for a selection of statements; whiskers show one standard deviation. Source: Table 3 of the paper.

Show as table

System	Average agreement (Likert scale)
A1 Consolidates theoretical AI concepts	4.34 ± 0.73
A4 Teaches new programming concepts	3.63 ± 1.08
B1 Course assessment before the experience	3.93 ± 0.87
B2 Course assessment after the experience	4.22 ± 0.77
C5 Teaches cooperation and teamwork	4.30 ± 0.95
D1 Competing improves team collaboration	4.23 ± 0.89
D2 Competing improves collaboration with other teams	2.85 ± 1.19
D3 Competing promotes motivation and interest	4.46 ± 0.83
D4 Participation is linked to winning the first prize	2.45 ± 1.21

Students agree that the project consolidates AI concepts and teaches problem solving, but less so programming. The course assessment rises from 3.93 to 4.22 after the experience, and all five skill statements score above 4. Competition motivates students and strengthens collaboration inside teams, but not between teams. For almost all statements in groups A to D the t-test gives a p-value below 0.001; the exception is D2, with a p-value of 0.106.

The pandemic did not change the overall picture. Comparing answers before and during COVID-19, only three statements differ significantly (p < 0.05), all about organization and collaboration.

Statement	Before COVID-19	During COVID-19	p-value
C4 Knowledge on work organization	4.40 ± 0.76	4.00 ± 1.04	0.003
C5 Knowledge on cooperation and teamwork	4.45 ± 0.82	4.16 ± 1.04	0.035
D2 Collaboration with other teams	3.05 ± 1.15	2.67 ± 1.21	0.038

Average Likert score ± standard deviation (Table 4 of the paper).

The open question collects suggestions for improvement. The most frequent ones concern technology, such as the server code (13% of answers), timing and deadlines (11%), and the explanation of the task (10%). Another 10% of answers only express appreciation for the experience.

Takeaways