Macau Strategy RL Arena
High-performance Polish Makao card game with Pure Zero-Sum PPO, Information Set MCTS, and unified simulation engine on Apple Silicon M1.
Macau Strategy RL Arena
tips_and_updates The 5 Morning Sparks
Any games that is non-digital firstAI driven Habitica
Habitica is an app I tried using in the past. It allows completing habit based real life tasks to improve the in-game character. You can set your own goals, tasks, etc. Then you can join skill guilds and spend coin on items in-game to flex the real life habit progress. What was missing is the a more seamless experience that an AI could provide.
Board game instruction chat
Sometimes the rules are unclear on the first read, sometimes we argue over the rules. What if there was an authority with perfect knowledge of the rules and references to show them? I imagine some RAG based system so the question can be vague-ish. Focus on text rules, not the images.
Macau strategy RL
Macau is a card game I played a lot as a kid. It has simple enough rules, but high complexity that is hard to comprehend. I want to try find a good strategy for playing the game. There are some guides online, but Reinforcement Learning could probably optimise it. A bot I can test myself against would be great.
Board game themer
Some board games get multiple themes, some board games have none. Can we get a system that takes in the board game rules and helps ideate over the thematic update to the mechanics and game elements? Both visually, and flavorwise in the rules text and box.
City guide trail
Real Life games are kind of a thing now. There is interaction with various people along the way, which would be the hard part, but the most important is to make the trip feel worth it. If someone is being yanked around random places, it feels bad. If it's effort put into seeing and doing cool stuff, it's great. Puzzles, historical facts, strong overarching themes welcome.
How It Was Built With AI in 4.0 Hours
Step 1
Clarified Makao rules, dropped the queen rule, and spun up the repo and web app with training view
Step 2
Playtested the game and fixed core rules like Ace suit selection and multi-card combos
Step 3
Discovered RL agent was degrading against the heuristic bot by learning bad card-dumping habits
Step 4
Added an MCTS agent with belief-state sampling for imperfect information
Step 5
Fixed broken best-of-5 stats in the UI, added MCTS parameter controls, and cleaned up the README
Step 6
Replaced broken Q-learning with pure zero-sum PPO with terminal-only rewards
Step 7
Fixed the missing voluntary draw mechanic ('dobieram') so bots can hold onto defensive shields
Step 8
Refactored into one file per agent, centralized configs, and unified simulation logic across the project
What Rocked
- ✓The bots were learning straight away, and MCTS suggestion was really good
What Broke & Fixed
- ✗The rules between the training and simulation not synching. The biggest issue was programs degrading many times when improvements were suggested with AI thinking it did a great job. It also made some bad assumptions and tried to convince me that I can't be right
smart_toy AI Interaction Summary
Genesis Prompt: "/spec-implementation" Interaction Turns: 31 human turns across 2189 execution steps Key Guidance Turns: • Turn 2: Remove the queen rule. Also, verify the rules about playing cards. A player should be able to play any number of cards of the s... • Turn 3: can we also get a training view in the Web UI • Turn 4: the 5-ideas app only needs to create a link to the Git Repo and a screenshot or two of the actual UI (use Playwright) • Turn 5: yes • Turn 6: the readme is empty • Turn 7: I can see issues with the rules already after playing. Ace doesn't work. It doesn't let me change suit. The 3 was played on 2 o... • Turn 8: Does the algorithm play against itself? • Turn 9: 60% win rate against heuristic agent is pretty bad. It should be close to 96% on best of 5, which 60% would not indicate • Turn 10: I can see in training that the agent win % against the heuristic steadily decreases over steps. It looks suspiciously like the ... • Turn 11: dumping a multi-card combo is not perfect arithmetic, it's a strategy that is actually not that good. No player misses a defenc... • Turn 12: do this, and reason also about whether this is a complete set of state that can be tracked. I assume the stack is being tracked... • Turn 13: so it got worse • Turn 14: The argument that because there are games that can not be won means that 10% edge is reasonable is a big leap in logic. If you ... • Turn 15: oh and 9 vs 3 cards is not a losing game if the 3 card player can't play a counter on a follow up or has to draw • Turn 16: yes fix this • Turn 17: so it's worse than originally still • Turn 18: yeah lets do that. Don't worry about how long it takes, as long as it's within a realm of viable execution on the M1 chip • Turn 19: the best of 5 on the webpage is broken. Also the table with training results is very confusing. It shows very low win% and erra... • Turn 20: ★ PROJECTED BEST-OF-5 RATE undefined% also against heuristic the training data is quite significantly off the 100 match simulat... • Turn 21: the readme is missing docs for running the new training. Overall make it user friendly no emojis, just clear explanation of wha... • Turn 22: can I control the parameters on the MCTs? • Turn 23: It seems to me like the Q-learning one is still misaligned. I like the speed, so I want to get it better. It currently loses qu... • Turn 24: resume • Turn 25: can we make it work in general, lets say when I train it for 10000 steps? I want to be able to see the plateau • Turn 26: are you sure you are correct and that we aren't optimising for plateau, rather than fixing a well built rl task? The win rate a... • Turn 27: Approach B is more interesting • Turn 28: I also found that even the MCTS program doesn't consider drawing when it has an option to play a card • Turn 29: Can you refactor the code into a better structure, commonize the game and simulation logic rules etc. I think there is a misali... • Turn 30: /capture-implementation • Turn 31: The steps should be based on the whole interaction. Any turn of my instructions should be included and condensed into 8 steps 2... Tools Executed: view_file (415), run_command (297), replace_file_content (125), manage_task (75), write_to_file (42), grep_search (33) Files Modified: README.md, RULES.md, SKILL.md, __init__.py, agents.py, analyze_losses.py, app.py, base.py