第７回　エンターテイメントと認知科学シンポジウムプログラム
The 7th Entertainment and Cognitive Science Symposium Program

　※参加無料、事前申込み不要

３月１８日（月）

オープニング：９：００−９：１０　　　伊藤毅志（電通大）

一般講演１：　９：１０−１０：３５　　座長：松原仁（はこだて未来大）
１−１．評価関数の機械学習を用いた自然な弱さを持った思考ゲームAIの作成（S)
　　　仲道隆史、伊藤毅志（電通大）
１−２．将棋の局面からの進行度の推定（S)
　　　小林健洋、小谷善行（農工大）
１−３．将棋における棋風を模倣するAIの構築（L)
　　　澤宣成、伊藤毅志（電通大）
１−４．FPSゲームにおける人間的AIの構築（S)
　　　森大道、伊藤毅志（電通大）
１−５．偶然手番感度によるゲームの特徴分析（S)
　　　西野順二、西野哲朗（電通大）

一般講演２：１０：４５−１１：５０ 　　座長：中村貞吾（九工大）
２−１．コンピュータ囲碁におけるモンテカルロ木探索への変形パターンの適用(S)
　　　下川和也、村松正和（電通大）
２−２．大貧民におけるゲーム中着手を反映させたプレイアウトによるモンテカルロ法（L)
　　　伊藤祥平、但馬康宏、菊井玄一郎（岡山県立大）
２−３．カーリングの戦略を支援するためのシステムの作成（L)
　　　北清勇磨、伊藤毅志（電通大）


特別企画：「アドバンスド将棋は最強コンピュータ将棋に勝てるか？」
　　協力：マイナビ、株式会社マグノリア

スケジュール：
　１３：００−１３：１０
　　　開会（対局者紹介、プログラム紹介）
　１３：１０−１５：１０
　　　第１局：「篠田 正人 氏 with Bonanza5.1」 VS GPS将棋
　　　　　　解説：古作　登　
　１５：４０−１７：４０
　　　第２局： 「古作 登 氏 with 激指12」 VS GPS将棋
　　　　　　解説：篠田　正人　

懇親会、表彰式　１８：００〜
　ハルモニアにて（こちらは、有料です）
　申し込みはこちら。

３月１９日（月）

一般講演３：　９：３０−１０：１０　　座長：角田博保（電通大）
３−１．Kinectを用いたジャグリングの技判定システム（L)
　　　長岡俊男、伊藤毅志（電通大）
３−２．Kinectによる手書き認識を用いた学習支援システムの開発（S）
　　　赤澤紀子、武井優樹、中山泰一、角田博保（電通大）

一般講演４：１０：２０−１１：５０　　座長：保木邦仁（電通大）
４−１．モンテカルロ囲碁における動的コミの多重適用（L)
　　　伊藤秀将、小谷善行（農工大）
４−２．NMFによる囲碁の棋譜分類(L)
　　　丸田要、永井秀利、中村貞吾（九工大）
４−３．囲碁における合議の為の多様なプレイヤ生成と予測市場メカニズム(L)
　　　立石真也、池田心（北陸先端大）
４−４．教師データ重要度の機械学習による合議囲碁プログラム（S)
　　　野中翔平、中村貞吾（九工大）



特別セッション：Cutting-Edge Computer-Go Programs １　
１３：００−１４：５０　座長：村松正和（電通大）

13:00-13:45 David Fotland
"Go Knowledge in The Many Faces of Go"

Many Faces of Go still contains code that was written 30 years ago. This talk will start with a brief history of computer go, including the strong programs, the competitions, and the techniques used. Before 2006 the strong programs depended on local tactical search, expert-knowledge pattern databases, and very limited full board search. Today, all strong programs use Monte Carlo Tree search. They use a mix of expert knowledge and knowledge from machine learning to guide the search. Today's Many Faces of Go also uses Monte Carlo Tree Search, but uses the knowledge from the old program to bias the tree search. Go knowledge is also used to control the probability distribution in the playouts. This talk will describe how Many Faces uses go knowledge, with examples.


14:00-14:45 Remi Coulom
"CLOP: Confident Local Optimization for Noisy Black-Box Parameter Tuning

Artificial intelligence in games often leads to the problem of　parameter tuning. Some heuristics may have coefficients, and they　should be tuned to maximize the win rate of the program. A possible　approach is to build local quadratic models of the win rate as function of program parameters. Many local regression algorithms have　already been proposed for this task, but they are usually not robust　enough to deal automatically and efficiently with very noisy outputs　and non-negative Hessians. The CLOP principle, which stands for　Confident Local OPtimization, is a new approach to local regression　that overcomes all these problems in a simple and efficient way. CLOP　discards samples whose estimated value is confidently inferior to the　mean of all samples. Experiments demonstrate that, when the function　to be optimized is smooth, this method outperforms all other tested　algorithms.

特別セッション：Cutting-Edge Computer-Go Programs ２　
１５：００−１６：５０　座長：村松正和（電通大）

15:00-15:45 Petr Baudis
"Information Sharing in Monte Carlo Tree Search"

We believe that new information sharing techniques are crucial for further progress of Computer Go. We can describe information sharing as a way to collect and reuse more information than just the simulation outcomes in the MCTS algorithm, allowing for propagation of information between iterations and across the tree search and simulation phases. We can collect more data on move performance in simulations and important branches in the tree, and reuse this data in future simulations as well as to bias the tree descent. Our goal is to use information sharing as a way to tackle the horizon effect, allowing simulations to resolve various local situations based on previous performance of moves and move sequences in similar situations. If this is successful, programs will not end up in a hopeless situation when a misevaluation occurs, and it will be possible for MCTS programs to correctly resolve even situations where a precise long tactical sequence is required, another common weakness of the current programs. We will give an overview of established and successful MCTS information sharing approaches (RAVE, criticality, last-good-reply, etc.). We will also present our own recent research in this area (both published and unpublished, in particular goal-based local trees and liberty maps) and solicit feedback and discussion about these and other approaches in the audience.


16:00-16:45 Martin Muellar
"Move Quality in Monte Carlo Simulation: A Case Study using the Fuego Go Program"

Despite half a dozen years of intense research, designing effective simulation policies for Monte Carlo Tree search in Go is still considered something of a black art, and driven largely by trial and error. Important ideas that have evolved include pattern- and tactics-based playouts, simulation balancing, and several schemes to dynamically modify simulation policies online. In this study, we take an in-depth look at what happens when the Go program Fuego runs its playouts. We develop several methods to evaluate the quality of moves played in these simulations, and we evaluate the contribution of the different components of Fuego's playout policy. We study the distribution of both the number of blunders, or result-changing moves, and the absolute loss - in terms of number of points - for many variations of the Fuego playout policy. We use this study to identify an improvement to the Fuego default policy.


一般講演５：１７：００−１７：５５　　座長：西野順二（電通大）
５−１．多様なシナリオを学習・提示する街づくりAIの提案（S)
　　　山本大祐、池田心（北陸先端大）
５−２．SNSにおける不適切な投稿文の修正を行うインタフェースの提案（L)
　　　高谷真弓、角田博保、赤池英夫（電通大）
５−３．心の社会理論に基づく多人数不完全情報ゲームにおける意思決定手法（S)
　　　藤村真理、西野哲朗（電通大）