AI Social Agents

点击翻译按钮获取中文摘要

Emote Games are developing a social networking platform upon whichgames will be played by a network of people. Such networks areenhanced by the presence of Artificial Intelligence (AI) members,which can direct human members to perform certain activities, offertutorial support, set up subgroups, etc. State of the art AI socialagents are fairly unimpressive - they are reactive rather thanproactive, and they tend to react in scripted and largelyuninteresting ways. It is fair to say that such agents are viewed asutilitarian in nature and not particularly interesting to interactwith. There is therefore much room for improvement, and to becomemarket leaders, Emote Games want their AI agents to add greater valueto the society. However, building agents to act intelligently (andperhaps moderately creatively), is a non-trivial problem which willrely on the use of various AI methods, including constraint solving,multi-agent systems, machine learning, planning and natural languageprocessing.Emote will iteratively build increasingly smarter AI agents for theirnetwork, with functionality that transfers from game to game. Each newrelease of the agents will perform more tasks, and perform existingtasks in more intelligent ways. To do this, each new release will beinformed by systematic experimental testing, which will be carried outat Imperial College. The Combined Reasoning Group at Imperial(www.doc.ic.ac.uk/crg) has much experience of building integratedsystems that deliver the kind of multifaceted intelligent behaviourrequired for this project. At the top level, we will build amulti-agent system architecture based on the beliefs, desires andintentions (BDI) model of agent behaviour. BDI agents are ideallysuited for situations where they are embedded, social, reactive, andgoal directed. This is a partial match to our requirements, and wewill experiment with various extensions which may be more suitable formore pro-active situations. We intend to determine the bestcommunication and behaviour framework to control the interaction ofthe agents within the network.Each agent will be semi-autonomous and will, like any other networkmember, have tasks to perform with respect to the game context. Wewill experiment with various planning approaches to control theoverall execution of the task, which will involve appealing to variousAI methods. As an example, an AI agent might be required to broadcasta piece of news to the right kind of network members, receive feedbackfrom the network members, and learn to do the same kind of task betterin future. Within an overall plan of action, determining which membersof the network to inform will be solved using constraint satisfactiontechniques. The communication of the news and the parsing of feedbackcomments will involve natural language processing techniques. Finally,given positive and negative feedback from the network members, theagent will set up a machine learning problem to generate a classifierwhich can be used in future to refine the constraint-based search fornetwork members to communicate news to.Emote Games will build the overall social network and the AI agents in it. They will specify the kinds of activities that agents will undertake, and write simple static scripts to do this. Imperial will beresponsible for suggesting planning methods and AI techniques toundertake the tasks in a more believable, intelligent and engagingfashion. Moreover, Imperial will deliver experimental results whichdemonstrate which approach(es) are the best for particular tasks. Toperform these kinds of experiments, we plan to work in controlledsituations where we gain feedback from subjects about theirsatisfaction with their interactions with the AI agents. In caseswhere the performance of the agents is sub-optimal, Imperialresearchers will research novel AI methods to improve matters.