今天開始著手撲克牌機器人的策略核心引擎。其實前一陣子就花了挺多時間測試和思考後，發現這個問題遠比想像中複雜。起初，我被那些能打敗職業選手的AI機器人所吸引，想要往機器學習的方向前進。然而在深入研究後，意識到這條路對新手開發者來說難度太高 - 不論是運算資源還是技術門檻，都遠超出這個Project的範圍。
轉而開始探索大型語言模型的可能性，分別測試了Claude 3.5 Sonnect、DeepSeek-V3和OpenAI o1的模型。在最基礎的測試中(prompt上完全沒有做什麼優化)，OpenAI的表現出乎意料地好 - 它能夠從基本的撲克資訊，像是大小盲、玩家位置等資訊，推理出現在的情況，並給出合理的建議。相比之下，其他模型在理解遊戲狀態時出現了一些基礎性的誤解，而且傾向給出較為保守的策略建議，這在現代撲克環境中可能不太適用。
不過OpenAI模型也有其限制：花過久時間做出回應和超高成本都是需要考慮的實際問題。這讓我開始思考幾個可能的改進方向：
1. 優化prompt設計，讓模型能更快速準確地理解場景
2. 建立基礎策略數據庫（將pre-solve solution轉成spreadsheet之類的），讓AI有個參考基準
3. 考慮做模型微調，注入更多現代撲克策略的知識
現階段的測試雖然初步，但已經證實了語言模型在這類任務上的潛力。接下來幾天我會繼續探索這些方向，尋找一個能平衡效能、成本和實用性的解決方案。
測試的Prompt
=== Table State ===
HU No Limit Holdem poker table, Blinds $1/$0.5
Street: Preflop
Hero cards: ['Qd', '6d']
Community cards: []
Hero stack size: $99.50
Villain stack size: $99.00
Hero chip in front: $0.50
Villain chip in front: $1.00
Pot size: $1.50
Positions: {'SB': 'hero', 'BB': 'villain'}
Button positions: on Hero
================
Available Actions:
- FOLD
- RAISE options: [$2.5]
================
Your response should be in JSON format with the following structure:
{
"action": "FOLD/CALL/CHECK/RAISE/BET",
"amount": null or number (for raise/bet),
"reasoning": "brief explanation of the decision"
}
Use the default no limit holdem HU pre-flop strategy
DeepSeek-V3
{
"action": "FOLD",
"amount": null,
"reasoning": "In heads-up No Limit Hold'em, Qd6d is a marginal hand from the small blind. Folding is the standard play here to avoid playing a weak hand out of position against the big blind's range, which is likely to be stronger. This helps conserve the stack for better opportunities."
}
Cluade 3.5 Sonnet
{
"action": "FOLD",
"amount": null,
"reasoning": "Q6s is a weak starting hand in heads-up play. From SB, facing a BB complete, we should fold this hand as it has poor playability postflop and low equity against villain's range."
}
OpenAI o1
{
"action": "RAISE",
"amount": 2.5,
"reasoning": "In a standard heads-up strategy, Q6 suited is strong enough to open-raise from the small blind. It has reasonable post-flop playability and fits into a typical wide SB opening range."
}
