Home » Recreation » Archive by category "Game" (Page 2)

AlphaGo Zero 的計算量

AlphaGo Zero 論文裡有提到,用同樣的硬體 (4 TPU) 可以用 89:11 碾壓 AlphaGo Master (今年年初與柯潔下的那個版本),主要是得力於更高品質的 neural network 以及更強的選擇能力 (後面這塊應該是將兩個 nerual network 簡化為一後的好處):

This neural network improves the strength of the tree search, resulting in higher quality move selection and stronger self-play in the next iteration.

那麼對應的問題就會冒出來了,究竟 DeepMind 花了多少時間才能訓練出這個新的 nerual network?結果吳毅成教授在 Facebook 上先估算出來了:

這邊的 TPU 對 GPU 的推估應該是基於當時 Google 在說明 TPU 的部份「An in-depth look at Google’s first Tensor Processing Unit (TPU)」:

In short, we found that the TPU delivered 15–30X higher performance and 30–80X higher performance-per-watt than contemporary CPUs and GPUs.

用 GPU 大約是 12K 顆,反推回 TPU 大約也是千顆這個數量左右。而這個數量以目前已經將 TPU 商用化的 Google 來看應該是很輕鬆,只能說有錢真好 XD:

1. 從另外一個角度看, DeepMind 僅40天就可以把 40-block 版本練起來, 換算一下, DeepMind 等於用了約12000顆 1080 Ti.

Windows 將引入 TruePlay,推出作弊偵測的 API

在「Windows now includes gaming cheat detection at the system level」這邊看到微軟將會引入 TruePlay (然後跟 Sonostrueplay 衝名 XDDD) 作弊偵測機制。

很明顯的會有隱私問題,而也很明顯的微軟說不會有隱私問題:

As Microsoft notes, "to protect customer privacy, no data is shared or transmitted until permission is granted," and no information is sent until "processing has determined cheating is likely to have occurred."

這不是把人當傻子嗎,遊戲一開始就會要求你同意才能玩啊,所以資料一定會送出的啊... 而且 TruePlay 變成作業系統的標準配備後,作弊程式就會找 workaround 才會推出 :o

Facebook 也參戰參與 StarCraft 的 AI 測試,只是成績不太好...

Facebook 也參與了 StarCraft 的 AI 測試,不過成績不太好:「Facebook Quietly Enters StarCraft War for AI Bots, and Loses」。

不過跟 DeepMind 投入的項目不太一樣,Facebook 投入的是 StarCraft,DeepMind 投入的是 StarCraft II...

比賽有二十八組,Facebook 拿下第六名,但前三名都只是獨立愛好者:

Final results released Sunday indicate Facebook still has a way to go: CherryPi finished sixth in a field of 28; the top three bots were all made by lone, hobbyist coders.

不太像認真要玩...

星海爭霸 II 官方的 AI Workshop

Blizzard 公佈了在十一月的月初將會舉辦星海二的 AI Workshop:「Announcing the StarCraft II AI Workshop」。

On November 3 and 4, Blizzard and DeepMind will co-host the StarCraft II AI Workshop at the Hilton Anaheim hotel, next to the Anaheim Convention Center.

官方 (包括 DeepMind 團隊) 也會針對 SC2LE (Starcraft II Learning Environment) 與 SC2API (StarCraft II API) 提供交流:

Engineers and researchers from Blizzard and DeepMind will also be on-hand to meet with attendees and answers questions about the SC2LE and SC2API.

然後時間會跟 BlizzCon 2017 重疊 (目前看起來是卡到最後兩天),票是不能通用的:

While this event takes place during BlizzCon 2017, it is considered a separate event and is not part of the official BlizzCon program – therefore BlizzCon badges will not grant access to the AI workshop. However, we will be providing a limited pool of shareable BlizzCon badges that attendees of the AI workshop can use to check out BlizzCon and catch the StarCraft II Global Finals for inspiration on how to build superior AIs!

接下來應該會有不少消息出來... DeepMind 團隊的開發進度有可以跟頂尖選手競賽嗎?

用 Xbox 控制器操控潛艇

拿民用品當作軍事器材不是第一次了,不過拿來操作潛艇倒是蠻值得看一下的:「The U.S. Navy's most advanced submarines will soon be using Xbox controllers」。


在 lab 環境測試的畫面。

原因頗簡單,就是成本考量,而且很容易取得:

The company says the photonic mast handgrip and imaging control panel that cost about $38,000 can be replaced with an Xbox controller that typically costs less than $30.

不過有個問題啊,軍用品一般都可以抵抗 EMP,這種民用品應該不行吧 XDDD

歐盟對於盜版是否帶來傷害的研究

歐盟在 2014 年做了關於盜版與銷量的研究,結果一直被壓到最近才發表出來 (於是就大概可以猜到結論了...):「EU Piracy Report Suppression Raises Questions Over Transparency」。

“In general, the results do not show robust statistical evidence of displacement of sales by online copyright infringements,” the study notes.

甚至:

The study found that piracy had a slightly positive effect on the videogames industry, suggesting that those who play pirate games eventually become buyers of official content.

另外也描述了現有電影與 TV-series 定價策略偏高:

“Overall, the analysis indicates that for films and TV-series current prices are higher than 80 per cent of the illegal downloaders and streamers are willing to pay,” the study notes.

難怪被壓著...

AlphaGo 與柯潔下的三盤棋,包括了雙方的講解

Aja Huang (黃士傑) 的 Facebook 上看到 AlphaGo 與柯潔在烏鎮下的三盤棋的講解,這次的講解除了找柯潔與樊麾以外,更重要的是直接拿了與當時相同配備的 AlphaGo 出來使用 (只要柯潔想要看某些變化 AlphaGo 會怎麼應對,樊麾都會輸入進去跑模擬):

需要說明的是,視頻中參與覆盤的AlphaGo與烏鎮峰會比賽的版本和硬體配置(搭配4個TPU的單機版)完全一樣。希望大家享受這三盤精彩的對局,也能從這三個視頻的覆盤研究中有所收穫。祝大家觀影愉快。

PS4 下載速度很慢的原因

在「Why PS4 downloads are so slow」這篇作者花了不少力氣找出原因,發現 PS4 下載速度很慢是故意的... 另外討論了在什麼情況下會變慢,以及要怎麼避免的方式。

懶得看的人可以直接看 Conculsions 那段,主要的原因是 PS4 會因為背景程式而調整 TCP window size (就算背景程式在 idle 也會影響到下載的 TCP window size),進而影響速度:

If any applications are running, the PS4 appears to change the settings for PSN store downloads, artificially restricting their speed. Closing the other applications will remove the limit.

用 TCP window size 來調整速度也算是頗有「創意」的方法...

Anyway,遇到時的解決方法就是把所有在跑的程式都完整關掉,再下載就會正常多了...

關於圍棋貼目的問題...

前陣子 AlphaGo 大獲全勝後放出了五十盤自戰棋譜 (兩台 AlphaGo 自己下),其實有件事情有點出乎大家意料,而在圍棋界被一直討論。就是在這五十盤裡,黑棋與白棋的勝率比是 12:38 (中國規則,黑棋貼 7.5 目的情況),明顯白棋有強大的優勢。

這個 7.5 目指的是,由於黑棋先下 (先手優勢),所以圍的地會比較多,為了彌補白棋後下的這個缺點,一般都會設計「貼目」這個規則。

交大資工的 CGI 團隊在上個月月底發了一篇論文 (參考「CGOS Whole Period Ratings for 19x19 Board」這邊的記錄,在有參加 CGOS 的團隊裡只輸新版的 Zen),討論 value network 的新想法:「Multi-Labelled Value Networks for Computer Go」。

他們對貼目的數量做了分析:

For the training data, we label on output 𝑣𝑘 as follows. For each self-play game, first calculate territory difference 𝑛 at the end of the game. Then, based on the Chinese rule, label 1 (win) on 𝑣𝑘 for all 𝑘 < 𝑛, and -1 (lose) for all 𝑘 > 𝑛. (Note that the draw case 𝑘 = 𝑛 is ignored in this paper since the komi is not an integer normally.) For example, if black occupies 7 more points of territory than white, the 𝑘-komi game is considered a win for all 𝑘 < 7, and a loss for all 𝑘 > 7. Thus, in this case, a 7.5-komi game is a loss, and a 6.5-komi or 0.5-komi game is a win.

這個研究完全顛覆了目前職業棋手一般的理解。目前的理解是,貼 5.5 目是黑棋優勢,貼 7.5 目是白棋優勢 (所謂的大貼目時代)。

接下來應該會有更多的研究出來,圍棋界會不會反思貼目規則呢...

Archives