Research Areas
|
|
|
|
FC Portugal Research AreasThis page contains a brief description of some of our research topics related to RoboSoccer and outline our approach in building a RoboSoccer team. It also contains a list of possible applications of the methodologies we developed for the simulated RoboSoccer domain. Until the moment, our research has been mainly focused on Multi-Agent Cooperation and Communication issues:
(Our research about RoboCup is still in the beginning. So, the research has been very much RoboSoccer focused. However, we tried to propose methodologies applicable to other domains. Unfortunately, those methodologies, at the moment are only tested in RoboSoccer.) Research Topic 1: Coordination of Teams of Homogeneous / Heterogeneous Agents in Adversarial environmentsProposal 1: Distinction Between Strategic and Active Situations Description: An active situation, is a circumstance in which an agent is needed to perform some type of active behavior soon. It must abandon its strategic position because the situation is critical in some way. For RoboSoccer this includes ball possession and ball recovery actions. A strategic situation is a circumstance in which an agent must keep a strategic behavior. Nothing special to do, so it is better to keep the strategic position that is supposed to be useful for the team. Proposal 2: Situation Based Strategic Positioning (SBSP) Description: This mechanism uses the situation information at several resolution levels to select the best strategic positioning for each one of the agents in the team. This strategic positioning depends on the agent positioning, the current tactic and formation of the team, the agent role (player type) inside the tactic and the current situation (described by multi-resolution information). Research Directions: Extend this framework to the
coordination of heterogeneous agents Description: Agents exchange their positioning and role (type of behavior) inside the team. Exchanges are based on utility functions based on positioning importance, cost to assume the new positioning, role adequacy, agent resources, etc. Research Directions: We are working on the concept of dynamic covering. This is a complementary mechanism to DPRE in which agents do not exchange positioning and role but only cover important unguarded positions temporally. Research Topic 2: Concept of Strategy for a Competition against other Team with Opposite GoalsProposal 4: STRATEGY - Formalization of what is a strategy for a competition against other team. Description: A strategy is composed by a set of tactics, agent types (roles), opponent modeling strategies, teammate modeling strategies, communication strategies and intelligent perception strategies. Each tactic includes formations that are used according to the situation (one at each time) and preset plans used in specific situations. In each formation, agents have different positioning and roles (player types). In strategic situations, agents keep their strategic behavior (defined before the competition). Only in active situations (defined by critical situation rules), agents assume an active behavior and abandon their strategic behavior. Research Directions: Create a completely general formalization to the definition of a strategy in a competition against other team. Research Topic 3: Communication in PTS Domains for Coordinating Teams of Autonomous Agents.Proposal 5: Balancing Active Communication and Strategic Communication. Description: Communication is needed both to coordinate actions and to ensure that the team strategy is being followed. We propose a communication mechanism that enables both strategic and active coordination using an unreliable channel with very low bandwidth. Research Directions: Improve this communication mechanism enabling to achieve better coordination in more constrained environments Proposal 6: ADVCOM - Intelligent Communication based on a Communicated World State. Description: A communicated world state is created only based on what is heard by the agent (without using vision or other perception). Based on this communicated world state, the agent may evaluate the information its personal world state has to the other agents. Research Directions: Study the relations between the world state update model, intelligent communication strategy and intelligent perception strategy. Research Topic 4: Intelligent Perception and Sensor FusionProposal 7: SLM - Strategic Looking Mechanism. Description: The interest of a given part of the world state to an agent is different according to the situation. SLM decides the agent looking direction according to the situation and with the confidence it has on the positions and velocities of all the objects. For each of the possible looking directions, a utility measure is calculated that estimates the usefulness for the agent to look on that direction and the best is selected. Research Directions: Improve the use of situations in SLM. Study the relations between the world state update model, intelligent communication strategy and intelligent perception strategy. Research Topic 5: Optimization Techniques applied to RoboSoccerProposal 8: Modeling low-level skills as optimization problems. Description: Instead of using an analytical model or use learning for the low-level skill, we propose to model low-level skills problems as optimization problems and solve them online. Research Directions: Model other problems (dribbling, holding the ball, intercepting, etc) also like optimization problems. Proposal 9: Optimization Kick. Description: Definition of an optimization model for the kicking problem. Online search algorithm divided in two phases: random search algorithm and neighborhood search algorithm. Research Directions: Use meta-heuristics (simulated annealing, tabu search, genetic algorithms, etc.) for solving this problem. Compare the results and efficiency of the algorithms. Research Topic 6: Modeling Soccer Knowledge for Integration in Intelligent AgentsProposal 10: Ball Possession Decision Module for RoboSoccer. Description: For playing soccer, agents have five different ball possession high-level actions: shoot, pass, forward, dribble and hold. We propose evaluation measures for each of these actions and a method to compare them. Research Directions: Gather more ball possession soccer knowledge from professional soccer coaches and integrate it. Proposal 11: Ball Recovery Decision Module for RoboSoccer. Description: For playing soccer, agents have eight different ball recovery high-level actions. We propose rules for deciding which is the best recovery action and when it should be used. The Marking Action is one of the most important and we propose a method base on marking utilities and teammate modeling to evaluate its usefulness. Research Directions: Gather more ball recovery soccer knowledge from professional soccer coaches and integrate it. Research Topic 7: Agent ArchitecturesProposal 12: Agent architecture for creating RoboSoccer agents. Description: We propose a very simple agent architecture for creating RoboSoccer agents. It build on top of the usual agent architecture, using a multi-level world state model and information about the team strategy (tactics, formations, player types, situations, communication strategy, communication strategy and opponent modeling strategy) that enable a somewhat complex high-level decision module to operate. Research Directions: Generalize this architecture to enable the agents to operate in other competitive domains. Research Topic 8: Visual Debugging and Analysis of Intelligent AgentsProposal 13: Visual Debugger. Description: Training, testing and tuning a soccer team is very important. So, we are researching the best way to show what the agents see, hear, feel, think and do in each game. Offline and human coaches may then use this information to improve the team. Our visual debugger, shows an offline graphical view of players’ minds at several different levels of abstraction that enables to do this. Research Directions: Study the best way to show this information to the human coach (using professional coaches knowledge). Proposal 14: Offline Client. Description: Low-Level Debugging Tool are very important and in any simulation, the ability to repeat offline the execution of an agent may be crucial in debugging. We propose the use of an offline client that enables this and the use of a debugger (like gdb) to examine variables and set breakpoints. Research Topic 9: Creating Accurate World States for Intelligent AgentsProposal 15: Multi-Level World State. Description: It is very important to have an accurate world model to decide the best action to execute. We propose the use of a multi-level world state as a way to reduce the complexity in the world state update process and in the use of this world state to select the most appropriate action. Research Directions: Improve the definition of this multi-level world state model and study methods for updating the different abstraction levels. Proposal 16: World State Error Analyzer. Description: We propose a tool that calculates mean error and standard deviation of the players world state using different weights. This enables the evaluation of intelligent perception and communication and world state update strategies. Research Directions: Use this tool to study the relations between the world state update model, intelligent communication strategy and intelligent perception strategy. Research Topic 10: Opponent Modeling in Adversarial EnvironmentsProposal 17: Very simple opponent model for RoboSoccer teams with three abstraction levels. Description: We propose a very simple model of the opponent team with three abstraction levels (team, individual and low-level). Team model concerns collective behavior, while individual model concerns individual decision and low-level concerns low-level skills of the opponents. Research Directions: Defining ways of updating this model and use it effectively to change the team strategy in accordance. Research Topic 11: Multi-Agent Learning in Adversarial Environments.Research Topic 12: Balancing Centralized and Decentralized Coordination (effective use of the online coach).Research Topic 13: Multi-Agent Planning.Research Topic 14: Agent Communication Languages for RoboSoccer (players and coach).Research topics beginning to be explored now! Possible ApplicationsThe following scenarios are possible future applications for some of our research developments: A1) Simulated RoboSoccer Collaboration in ProjectsAre you Interested in future collaborations in projects along these lines starting next year? Please contact Luis Paulo Reis (LIACC - UP / UFP) or Nuno Lau (IEETA - UA) |
|
