Presentation on theme: "Annual Conference of ITA ACITA 2009 Chukwuemeka D. Emele, Timothy J. Norman, Frank GuerinSimon Parsons Computing Science Department, University of Aberdeen,"— Presentation transcript:
Annual Conference of ITA ACITA 2009 Chukwuemeka D. Emele, Timothy J. Norman, Frank GuerinSimon Parsons Computing Science Department, University of Aberdeen, UK Brooklyn College, City University of New York. USA. The Problem Coalition members often operate under constraints (e.g. policies), which are not necessarily public knowledge. Developing and resourcing coalition plans require an understanding of the policy and resource availability constraints. Tracking and reasoning about such information is non-trivial. Standard machine learning is not effective in disambiguating between policy and resource availability constraints. The Goal To investigate how software agents could employ argumentation for teasing out vital information from dialogue and using that to aid the learning of underlying constraints that others operate with. Military Relevance This research is focussed on how software agents can support humans in collaborative decision making by: Keeping track of who might have and be willing to provide the resources required for enacting a plan. Modeling the policies of others regarding resource use, information provision, etc. The Approach We built an experimental framework that combines argumentation and machine learning for identifying and modeling the policies of other agents in the domain. Hypothesis Allowing agents to exchange arguments during dialogue will mean that the proportion of correct policies learned during interaction will increase faster than when there is no exchange of arguments. The Framework For simplicity, we present a setup with one seeker and a number of providers. The seeker simulates an agent that is looking to resource some plans to execute a task and needs to find the best partner to collaborate with. Plans are resourced by convincing a provider to release some resources from its resource pool. The Simulation Each agent has two main layers, the communication layer and the planning and reasoning layer. The communication layer embodies the dialogue controller, which handles the communication with other agents. The planning and reasoning layer consists of three modules: the planner, the policy modeller, and the learner. References 1. Chukwuemeka D. Emele, Timothy J. Norman, Frank Guerin, and Simon Parsons. Argumentation-based agent support for learning policies in a coalition mission. In Proc. of the Third Annual Conference of the International Technology Alliance, Maryland, USA, 2009. 2. Gita Sukthankar and Katia Sycara. Analyzing Team Decision-Making in Tactical Scenarios. The Computer Journal, page bxp038, 2009. 3. N. Oren, T. J. Norman, and A. Preece, Loose lips sink ships: A heuristic for argumentation, In Proc. of the Third International Workshop on Argumentation in Multi-Agent Systems (ArgMAS), 2006, pp. 121–134. Dialogue Snippets Why did agent j say no to is request? 1) Could it be that there exist policy X that forbids j from providing R1 to agent i? 2) Could it be that R1 is not available? There is very little that we can learn from the dialogue. However, suppose agents are able to ask for and provide explanations as in the examples below then agent i can gather more evidence regarding why agent j did not provide R1. Experimental Setup Three agent support configurations were tested and the performance of the seeker was evaluated. The configurations include: Random Selection (RS): The seeker does not employ argumentation or machine learning techniques, rather it randomizes its choice of attributing the refusal to policy or resource availability constraints. Learning without Argumentation (LOA): The seeker applies the C4.5 decision tree learner to learn the providers policies. Learning with Argumentation (LWA): Here, argumentation is used to augment the C4.5 learner in learning the policies of others The Results The results show that using learning with argumentation (LWA) enabled the seeker to learn and build a more accurate model of the other agents policies and thereby increased the accuracy of predictions. The LWA approach constantly out-performs machine learning only. Future Directions To develop strategies for advising human decision makers on how a plan may be resourced and who to talk to on the basis of policy and resource availability constraints learned from previous interactions. To investigate the suggestion of alternative resources based on similarity metrics. Fig 1. Architecture for Argumentation-based Agent Support for Learning Policies Example of a Policy Coalition member X is permitted to release resource R to another coalition member Y if Ys affiliation is O and the resource R is to be deployed at location L for purpose P on day D. Example 2: i: Can I have R1? j: No. i: Why? j: Not permitted to release R1. Example 3: i: Can I have R1? j: No. i: Why? j: R1 is not available. Example 1: i: Can I have R1? j: No. Argumentation-based Agent Support for Learning Policies in a Coalition Mission
Your consent to our cookies if you continue to use this website.