Algorithmic Game Theory - Lecture4

Algorithmic Mechanism Design

上一講利用 Myerson's Lemma 闡述了給定在「好」的 allocation rule 下,該如何定價,使得 auction 具有 DSIC property; 而這講則要說明該如何 derive 出「好」的 allocation rule。

Knapsack Auction

上一講定義了 single-parameter environment ,其中的 allocation set $X$ 需滿足元素和 $ \leq k$ (in single-item, $k=1$)。這裡再將此推廣成更一般的形式。

  • bidder $i$ has public size $w_i$ and private valuation $v_i$
  • The allocation set $X$ contains 0-1 vectors subject to $\sum_{i=1}^n w_i , x_i \leq W$

我們想要找的 allocation 需能最大化 social surplus

\[ \vec{x}(\vec{b}) = \arg \max_{\vec{x}} \sum_{i=1}^{n} b_i \, x_i \]

若想說明其具備 DSIC property,我們可以先證明該 allocation rule 為 monotone ,再利用 Myerson's Lemma 得出其 payment rule

[Theorem]

\[ \text{If} \: \vec{x}(\cdotp,\vec{b}_{-i}) \: \text{is NOT monotome in } [0,b_{max}] \Rightarrow \exists \: \vec{x^{\prime}} \quad \text{s.t} \: \sum_{i=1}^{n} \: b_i \, x^{\prime}_i \: > \: \sum_{i=1}^{n} \: b_i \, x_i \]

$\underline{~Proof.}$

\[ \begin{aligned} & \text{For bidder} \, i, \text{we construct an example that} \, \beta_1 < \beta_2 = b_i \, \text{, but} \, \vec{x}(\beta_1) > \vec{x}(\beta_2) \\ & \text{we pick} \, \vec{x^{\prime}}(\beta_2) = \vec{x}(\beta_1) \, \text{and} \, \vec{x^{\prime}}(\beta_1) = \vec{x}(\beta_2) \\ & \text{It is easy to check that} \, \vec{x^{\prime}} \cdotp \vec{b} > \vec{x} \cdotp \vec{b} \quad \square \end{aligned} \]

但目前為止看起來非常完美,但真是如此嗎?別忘了我們希望整個 mechanism 能在 POLY time 內結束,然而,尋找 max allocation 的問題(等價於找到 Knapsack problem 的最佳解)卻是一個 NPC …QQ,也就是說,對一般的 k-item auction 而言,不存在多項式時間的 mechanism。所以類似於在 approximation algorithm 所學的,我們的 objective 如以下。

Relax optimal surplus constraint as little as possible, subject to DSIC and POLY-time constraint

利用 Myerson's Lemma,以下敘述也等價

Design POLY-time alg. and monotone allocation rule as close as possible to maximize the social surplus

至此,在 approximation alg. 中學到的當前 state-of-the-art 的演算法,可用於解決 AGT 的問題 (W/ 還不錯的 ratio ),但有件事情需要注意

The DSIC/monotone constraint that approx. mechanism should satisfy may cause additional surplus loss

因為我們需要保證 DSIC ,因此對某些 approx. algorithm ,我們可能需要犧牲 approx. ratio 去 maintain 這個性質。 (e.g black-box reduction)

以下定理證明用於 Knapsack problem 的 2-approx algorithm 所設計的 allocation rule 不會造成額外的 cost ,同時保持 monotone 的性質。

[Theorem] The greedy allocation rule for Knapsack auction is monotone

2-approx Algorithm:   Sort bidders according to $ \frac{b_i}{w_i}$ in decending order
Choose bidders in that order until constranit DOESN'T satisfy,
then pick $max(b_1 + b_2 + \cdots +b_i,$ $b\scriptstyle{i+1}$$)$

For every bidder $i$, $\exists$ critical bid $b^{\star}$ (depending on its $w_i$ and $b\scriptstyle{-i}$) s.t $x_i$ jump from 0 to 1 。

It is easy to check that such allocation rule for arbitary bidder $i$ is monotone 0-1 curve 。

 
comments powered by Disqus