• Deep Reinforcement Learning with Applications in Transportation ◦ 講義資料:https://outreach.didichuxing.com/tutorial/AAAI2019/ • On Explainable AI: From Theory to Motivation, Applications and Limitations • Plan, Activity and Intent Recognition (PAIR) • Behavior Analytics: Methods and Applications • Building Deep Learning Applications for Big Data Platforms ◦ 講演資料:https://jason-dai.github.io/aaai2019/ • New Frontiers of Automated Mechanism Design for Pricing and Auctions ◦ 講演資料:https://sites.google.com/view/amdtutorial/home • Federated Learning: User Privacy, Data Security and Confidentiality in Machine Learning ◦ 講演資料:https://www.fedai.org/#/conferences/link_aaai2019 本講演でも触れます チュートリアル AAAI 2019: Tutorials https://aaai.org/Conferences/AAAI-19/aaai19tutorials/
講演資料:https://www.nms.kcl.ac.uk/andrew.coles/PlanningCompetitionAAAISlides.html • Presenting a Paper ◦ 講演資料:https://wp.me/P3qAAw-76 • Planning and Scheduling Approaches for Urban Traffic Control ◦ 講演資料:https://helios.hud.ac.uk/scommv/storage/TutorialSlides.pdf • The Road to Industry • Adversarial Machine Learning ◦ 講演資料:https://aaai19adversarial.github.io/index.html#org • Deep Bayesian and Sequential Learning ◦ 講演資料:http://chien.cm.nctu.edu.tw/home/aaai-tutorial/ • Multi-Agent Pathfinding: Models, Solvers, and Systems ◦ 講演資料:http://ktiml.mff.cuni.cz/~bartak/AAAI2019/ • Neural Vector Representations beyond Words: Sentence and Document Embeddings ◦ 講演資料:http://gerard.demelo.org/teaching/embedding-tutorial/
◦ 講演資料:https://preferred.ai/aaai19-tutorial/ • End-to-end Goal-oriented Question Answering Systems ◦ 講演資料:https://www.slideshare.net/QiHe2/aaai-2019-tutorial-endtoend-goaloriented-question-answering-systems • Graph Representation Learning • Imagination Science: Beyond Data Science ◦ 講演資料:https://people.cs.umass.edu/~mahadeva/AAAI_2019_Tutorial/Welcome.html • Integrating Human Factors into AI for Fake News Prevention: Challenges and Opportunities • Knowledge-based Sequential Decision-Making under Uncertainty ◦ 講演資料:http://www.cs.binghamton.edu/~szhang/2019_aaai_tutorial/ • Human Identification at a Distance by Gait Recognition ◦ 講演資料:http://yushiqi.cn/research/aaai19-gait-recognition-tutorial
for Industrial Automation Competition (ARIAC) • Artificial Intelligence for Cyber Security (AICS) • Artificial Intelligence Safety • Dialog System Technology Challenge (DSTC7) • Engineering Dependable and Secure Machine Learning Systems • Games and Simulations for Artificial Intelligence • Health Intelligence • Knowledge Extraction from Games • Network Interpretability for Deep Learning • Plan, Activity, and Intent Recognition (PAIR) 2019 • Reasoning and Learning for Human-Machine Dialogues (DEEP-DIAL 2019) • Reasoning for Complex Question Answering • Recommender Systems Meet Natural Language Processing • Reinforcement Learning in Games • Reproducible AI 本講演でも触れます
• ここで紹介する事例 ◦ ゲーム理論や機械学習を使ったAI for Social Good事例 ▪ 警戒行動、野生動物保護、感染症拡大防止 ため 介入 AAAI 2019: Emerging Track https://aaai.org/Conferences/AAAI-19/aaai19emergingcall/ • AI and Multiagent Systems for Social Good (Milind Tambe; AAAI2019 Invited) • On the Inducibility of Stackelberg Equilibrium for Security Games (Guo et al.2019; AAAI2019) • Melding the Data-Decisions Pipeline: Decision-Focused Learning for Combinatorial Optimization (Wilder et al.2019; AAAI2019) 紹介するAAAI2019/NeurIPS2018講演・論文
Massive Security Games (Kiekintveld et al.2009; AAMAS2009) http://teamcore.usc.edu/kiekintveld/papers/2009/kjtpot-massive-security-games.pdf • Federal Air Marshal Service ◦ 同様 アプローチ 、航空機に対する警備行動でも適用可能 ◦ 組み合わせが1041に及ぶ警備行動を、部分問題を拡張することで解いた GUARDS and PROTECT: Next Generation Applications of Security Games (An et al.2011) http://teamcore.usc.edu/people/marecki/sigecom.pdf • GUARDS and PROTECT ◦ 船舶 警備に適用した事例(現在 様々な国と地域で導入) ◦ ターゲティングスケジュール 探索次元を削る工夫をしている ◦ 当時 人間によるスケジューリングと比べ350% 利得改善
多い ▪ 経済活動、交渉、安全保障行動、ポーカー・麻雀など ゲーム、… ◦ 完全情報ゲーム(チェッカー・オセロ・チェス・将棋・囲碁)で 人間を超える方策が獲得できている • Deep Counterfactual Regret Minimization (Brown et al. 2018; NeurIPS2018) • Solving Imperfect-Information Games via Discounted Regret Minimization (Brown&Sandholm2018; AAAI2019) • New Results for Solving Imperfect Information Games (AAAI2019 Invited Talk) 紹介するAAAI2019/NeurIPS2018講演・論文
AI Labによる 交通に関連した深層強化学習事例 紹介 講義資料:https://outreach.didichuxing.com/tutorial/AAAI2019/ • Deep Reinforcement Learning with Applications in Transportation (AAAI2018 Workshop) • Deep Q-Learning Approaches to Dynamic Multi-Driver Dispatching and Repositioning. (Holler et al. 2018; NeurIPS2018 DRL WS) • Learning to Navigate in Cities Without a Map (Mirowski et al.2018; NeurIPS2018) 紹介するAAAI2019/NeurIPS2018講演・論文
DNNによって大規模都市へ 適用、転移学習 活用、汎化効果が期待 • Action Search ◦ 過去 軌跡を活用してデータ拡張することでスパースなグリッドに対応 Deep Reinforcement Learning with Knowledge Transfer for Online Rides Order Dispatching (Wang et al. 2018; ICDM2018) https://tonyzqin.files.wordpress.com/2018/09/drl_tl_dispatch_icdm_camera_ready.pdf
全車両 経過時間等を(負 )報酬として信号切り替えタイミングを制御 Deep Reinforcement Learning for Traffic Light Control in Vehicular Networks (Liang et al.2018) https://arxiv.org/abs/1803.11115
◦ 実際に変えられる信号 ステータス Using a Deep Reinforcement Learning Agent for Traffic Signal Control (Genders & Razavi 2016) https://arxiv.org/abs/1611.01142 レーン・グリッドごと 車両数 レーン・グリッドごと 平均車両速度
◦ Google Street View 画像を元にゴール(緯度経度)まで移動するタスク ◦ 画像認識にCNN、不完全観測に対してLSTMを活用(迷路タスクと同様) ◦ 方策獲得に 分散強化学習(IMPALA)を活用 ◦ カリキュラム学習 採用 ◦ 都市間 転移学習にも成功 Learning to Navigate in Cities Without a Map (Mirowski et al.2018; NeurIPS2018) https://arxiv.org/abs/1804.00168 フィーチャーサイト(動画あり) https://sites.google.com/view/streetlearn