Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Artificial Intelligence Laboratory Boi Faltings Artificial Intelligence Artificial Intelligence • People are good in the natural world: • irregular shapes • reactive • Modern world is artificial: • precision machinery • schedules and interdependencies • AI: helps people work in the artificial world My Motivation • Use computers to solve a problem that people could not solve otherwise. Sharing Networks Sharing Networks • Routing demands in circuit-switched communication networks. Sharing Networks • Routing demands in circuit-switched communication networks. Sharing Networks • Routing demands in circuit-switched communication networks. • Shortest path in blocking island graph rather than network => 40% more capacity. Iconomic Systems V1 Iconomic Systems V1 Iconomic Systems V1 • Goal: market Icononet = routing tool that increases capacity by 40% Iconomic Systems V1 • Goal: market Icononet = routing tool that increases capacity by 40% • Problem: nobody wanted to save capacity. Recommendation systems Infer preferences from past behavior. Collaborative filtering: use preferences of like-minded users. Problems: requires too much information !! ! (30-40 rated items per user) slow to incorporate new items. 7 Ontology Filtering Structure user preferences into an ontology. => better accuracy with less information (5 items) => new items can be placed in ontology and are immediately accessible. Chorafas award/patent. 8 • Startup company Prediggo commercializes technology. • Sold to about 40 web sites so far, including Moevenpick, MediaMarkt, Ricardo, etc. 9 Performance on actual web sites • Baskets impacted + 5% to 32% • Sales facilitated + 5% to 25% • Conversion rate + 10% to 30% • Long Tail items + 30% to 95% 10 News recommendation • Challenges: • short traces • items change all the time • avoid redundant recommendations • new technique: context trees Current Work • Human-level AI • Social Computing Human-Level AI • Can we match human performance on common sense tasks like natural language processing, vision, etc.? • Recent advances (e.g. Watson) make it seem so. • Everyone expects big things from big data... Limitations of Big Data • Learning from large corpora has hit a dead end: only marginal improvements for huge computation costs. • Nice feature of commonsense knowledge: do not need experts, everybody has it! • New direction: crowdsource knowledge acquisition Crowdsourcing Platforms Games Practical Impact • Analysis of social network interactions • Recommendations from review texts Awesome pool! 9$:(;%*'401'2342' Nice, clean hotel 567%'41'2338' Simply filthy … -./'01'2344' !"#$% &'(")'' *%+#%,&' =(("' ?($.:#(7' @((A' +#%,' B"%.7"#7%&&' <7:%*%&:' =*(>"%' !"#$%&'()*&%' Dirty towels ' Good location ' D+%' E(;' @%$(AA%7).:#(7 C' Social Computing • Intelligence based on learning from data gathered from individuals. • Quality of data is crucial! Great Leap Forward (China, 1958-60) • Progress measured by grain/steel production figures. • Village chiefs pressured to report inflated grain production (up to 10 times). • Result: grain was exported in spite of shortage, at least 30 million people died. Subprime Mortgages (USA,2008) • Risk of financial products assessed by statistical models. • Debt bundled into large packages cut into slices with different priorities and ratings. • Banks reverse-engineered models and made risky debt look unrisky. • We know the rest... Big Data... • Big data can become dangerous! • We need to ensure its quality: • filtering • motivation of data providers • Focus on crowdsourcing scenarios. Asking the Crowd • Collect information through the internet: • opinion forums • product reviews Asking the Crowd • Collect information through the internet: • opinion forums • product reviews Community Sensing Community Sensing • Goal: Maintain a map of pollution in a city. Community Sensing • Goal: Maintain a map of pollution in a city. • Need to distribute sensors in many places. Community Sensing • Goal: Maintain a map of pollution in a city. • Need to distribute sensors in many places. • Community sensing: citizens operate sensors and get paid for the data. Community Sensing • Goal: Maintain a map of pollution in a city. • Need to distribute sensors in many places. • Community sensing: citizens operate sensors and get paid for the data. Truthfulness • Reporting values is a lot of work. • Agents only do it for ulterior motives: • promote themselves or their friends • punish their rivals • take revenge Truthfulness • Reporting values is a lot of work. • Agents only do it for ulterior motives: • promote themselves or their friends • punish their rivals • take revenge Truthfulness • Reporting values is a lot of work. • Agents only do it for ulterior motives: • promote themselves or their friends • punish their rivals • take revenge Example: music ratings ratings on Amazon ratings in controlled experiment [Hu, Pavlou & Zhang, 2006] Better Results • Reward people for their efforts: karma points, rankings, gifts, or money. • Scale rewards to encourage truthfulness: • payoff is a function of report and outcome. • truthful report maximizes expected reward. Opinion Polls Monitoring Service Level Agreements Current Approaches to QoS Monitoring QoS Monitoring using Feedback from Clients Monitoring by Proxy (expensive) Client Monitor Provider QoS estimates Reputation Mechanism Monitoring by Sampling (imprecise) $ Client QoS Provider Client Decentralized Monitoring (not trustworthy) Client (cheap, reliable and precise) Provider Provider Community Sensing • Sensors placed in private properties • Reward them for providing data (similar to solar energy) • Incentive scheme rewards accuracy. Other Work • Digital health: recommendations from wearable sensor data. • Multi-agent optimization. • Collaboration with privacy guarantees. • Cognitive information security.