水果交易游戏场景下：非暴力推测未知变量的机器学习方案可行性探讨

First, let's clarify the core problem here: you need to infer the unknown quantity of each fruit in the basket (let's call this vector Q = [q₁, q₂, ..., qₙ]) given:

• Daily price vectors P(t) = [p₁(t), p₂(t), ..., pₙ(t)]
• Daily total basket value V(t) = sum(qᵢ * pᵢ(t))
• The constraint that any single buy/sell transaction changes the total basket value by ≤10%

Your current brute-force approach is almost certainly inefficient (especially as the number of fruit types grows), so let's cover better alternatives and address the machine learning question.

Better Alternatives to Brute Force #

1. Linear Algebra & Integer Linear Programming (ILP) #

This is the most direct and efficient approach if you have enough data:

• Each day's price and total value gives you a linear equation: P(t) · Q = V(t). If you collect n linearly independent price vectors (where n is the number of fruit types), you can solve this system of equations exactly to get Q.
• If quantities are integers (which makes sense for a game), this becomes an integer linear programming problem. Tools like branch-and-bound or cutting-plane algorithms are way more efficient than brute-force enumeration, as they prune impossible combinations early.
• You can leverage the 10% transaction constraint to narrow down feasible Q values: after any transaction, the new total value V' must lie in [0.9*V_old, 1.1*V_old]. This translates to constraints on how much each qᵢ can change between rounds, which eliminates invalid quantity combinations upfront.

2. Bayesian Inference #

If you don't have enough data to solve the linear system yet (e.g., early game rounds), Bayesian inference lets you iteratively refine your guesses:

• Start with a prior distribution for Q (e.g., assume quantities are non-negative integers within a reasonable range for your game).
• For each new day's P(t) and V(t), update the posterior distribution of Q using Bayes' theorem. This narrows down the most likely quantity combinations over time.
• Incorporate the 10% transaction constraint into your likelihood function: discard any Q values that would require a transaction exceeding the 10% value change to reach the previous round's V(t).

Is Machine Learning Applicable? #

Absolutely, but it depends on your specific needs:

• Supervised Learning (Regression): If you have historical data (past price vectors, total values, and eventually known quantities), you can train a regression model (e.g., Lasso regression to handle sparse quantities, or random forests for non-linear price patterns) to predict Q. This works well if your game has noisy price data or a large number of fruit types.
• Online Learning: For real-time, round-by-round inference, use online linear regression models. These update their predictions incrementally as you get new P(t) and V(t) data, which fits perfectly with your game's turn-based structure.
• When to Skip ML: If you have a small number of fruit types and can collect enough independent price data, linear algebra/ILP will be faster and more accurate than ML—no need to overcomplicate things.

Key Takeaways #

• Ditch brute-force first: linear algebra/ILP is the gold standard when you have enough data, and Bayesian inference is ideal for early-game, data-scarce scenarios.
• ML is a great option if you're dealing with large-scale fruit sets, noisy data, or want to build a system that adapts to long-term price trends.

内容的提问来源于stack exchange，提问作者Lostsoul