Homework 3 🚀

Back to the AGI Course - 2026 Fall (AI5102)

The third homework is a individual term project that runs until the end of the semester. You will build an agent for the Expanded Othello AI Arena, discuss with your classmates along the way, and present and write what you learned.

Purpose of this assignment:

Most benchmarks tell your agent the rules and ask it to play well. This one does not. The agent is told the board and nothing else, and every rule it relies on has to be inferred from what it is allowed to do and what happens when it does it. That makes this an exercise in test-time adaptation under a hard constraint: there is no training phase and no test phase, every game counts from the first, and the opponent is another student’s agent that is adapting to you while you adapt to it. A policy that is excellent on the environments you practised on and rigid everywhere else will lose.

The Arena:

• Platform: Expanded Othello AI Arena (Click)
• Docs: Getting started (Click)
• Sign-up: log in with your GitHub account.

You will write an agent that plays Othello on boards that are not always the standard 8×8 game. Some stages are public and you can practice on them as much as you like. Other stages are private and only the staff can see them. A good agent should not only win on the stages you have seen, but also adapt to the ones you have not.

aig-15 is a spare account for staffs.

Timeline:

What to submit:

This is an individual project. (deadline is 11:59PM, Korea Time)

1. Milestones (Nov 9, Nov 23). Upload your current agent to the platform. Then leave one comment in the milestone thread on Ed Discussion, which the staff will open for each milestone. Your comment should include:
   • your username or agent name on the platform, so the staff can match it to the leaderboard,
   • your approach so far, with at least one number (e.g., win rate on the public stages),
   • something that failed or surprised you,
   • a reply to at least one classmate’s comment.

   Share ideas and results freely, but do not post your full code.

2. Final agent (Dec 6). Upload your final code to the platform. Your performance on the public and private stages is graded, and the staff will check that your submitted code runs within the time limits.
3. Presentation (Dec 14; 09:00-12:00). A live, in-person talk of 8 minutes, followed by a short Q/A, in a 10-minute slot (longer presentation is not allowed). Keep it to less than 8 slides: your approach, the evidence, what failed, and how your agent handles stages it has not seen. The technical details belong in your report. Slides link will later be included in your report too. Sandwiches and coffee will be provided.
4. Report (Dec 18). The format will be discussed and shared later.

You may use an AI assistant while building your agent. If you do, describe how you used it in your report.

Grading Criteria (60%):

All weights are percentages of your course grade, so HW3 adds up to 60%.

• Milestones are graded as Check Plus / Check / Check Minus. A milestone without an agent on the platform loses points.
• Leaderboard performance is graded in tiers (e.g., beating the baseline agents on public and private stages), not by raw rank.
• The code check confirms that your submitted code runs within the time limits.
• The presentation is graded on clarity, evidence, and what you learned, not on leaderboard rank.

Late submission will be graded according to the late policy