P. Stone and M. Veloso. Using decision tree confidence factors for multi-agent control. Autonomous Agents, 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....machine learning seemed more promising than non symbolic machine learning approaches like neural networks since the latter could result in a model that is di#cult for a human to comprehend. Amongst symbolic approaches, decision trees have often been used for agents learning about own decisions [32] (or for modeling others [8] in the presence of large amounts of data. However, unlike these approaches that use decision trees as a model of prediction of agent behavior in unseen cases, we use decision trees as a model to explain observed agent behavior. The key insight here is that we wish to ....

....In contrast, ISAAC is knowledge lean, but has more di#erent types of analysis and presentation techniques and is capable of analyzing any RoboCup team. Stone and Veloso have also used a decision tree to control some aspects of agents throughout an entire game, also using RoboCup as their domain [32]. However, this work pertains to execution of agents rather than analysis of agent teams, and since it is internal to the agent, their work has no means of presentation. There is of course great general interest in analysis of (human) Soccer games, e.g. see [30] Human soccer games are much more ....

Stone, P., Veloso, M.: Using Decision Tree Confidence Factors for Multi-agent Control. Proceedings of the International Conference on Autonomous Agents, 1998.

....is shown in Table 2.2. In the bottom layer the ball interception behavior has been learned using a neural network. The pass evaluation behavior in the second layer has been learned using the C4.5 decision tree algorithm (see [75] and uses the learned ball interception skill from the layer below [93, 97]. Subsequently, the pass selection behavior in the third layer has been learned using a new multi agent reinforcement learning method called TPOT RL with the pass evaluation skill Team Partitioned Opaque Transition Reinforcement Learning: this method can be used for maximizing long term ....

....references we also summarize the significant results of these teams in international competitions (top 10 finishes only) It is important to realize that the list is not exhaustive and only contains references and results up to and including the year 2000. Team References Roll of honour CMUnited [7, 90, 92, 93, 94, 95, 96, 97] 4th WC97, 1st WC98, 98, 99, 100, 102, 103, 104, 115] 1st 9th WC99, 3rd 4th WC00 Essex Wizards [36, 37, 39, 50, 51, 52, 53] 3rd WC99, 3rd EC00, 7th WC00 FC Portugal [56, 76, 77] 1st EC00, 1st WC00 Cyberoos [13, 69, 70, 71, 72, 73, 74] 3rd PR98, 4th EC00, 9th WC00 Karlsruhe Brainstormers [79, ....

P. Stone and M. Veloso. Using Decision Tree Confidence Factors for Multi-Agent Control. In H. Kitano, editor, RoboCup-97: Robot Soccer World Cup I, pages 99--111. Springer Verlag, Nov. 1998.

....dependencies Agent behavior: Set of states, events and transitions So far, we have established the neccesity for a meta agent and enumerated its responsibilties. The next section deals with the actual design of one such meta agent. A. Meta agent design The meta agent decides on a strategy [14]. It determines what behaviors of the agents would achieve this strategy and accordingly triggers those behaviors. By triggering only the behaviors appropriate to the current strategy, the meta agent also reduces behavior behavior interactions. This is also in accordance with the layered ....

Peter Stone and Manuela Veloso. Using decision tree confidence factors for multiagent control. In H.Kitano, editor, RoboCup-97: The First Robot World Cup. 1997.

.... can learn social rules through reinforcement (Mataric 1994) and they can learn to solve tasks that require tight coordination by using communication to share sensory data (Mataric 1997) Agents in a team can achieve coordination by learning confidence factors pertaining to their team actions (Stone Veloso 1997). Balch (1997) analyzes how agent team members can activate various behaviors depending on the reinforcement that is provided at team level or at agent level. In conflicting situations agents use learning about their preferences to build models that help them avoid conflict (Grecu Brown 1996) ....

P. Stone and M. Veloso. "Using Decision Tree Confidence Factors for Multi Agent Control," AAAI-97 Workshop in Multi-Agent Learning, edited by S. Sen, 1997.

....different. ISAAC performs something similar in its perturbation analysis; however, ISAAC focuses on an entire team, not just an individual, necessarily. Stone and Veloso have also used a decision tree to control some aspects of agents throughout an entire game, also using RoboCup as their domain [14]. However, this work pertains to execution of agents rather than analysis of agent teams, and since it is internal to the agent, their work has no means of presentation. 8. CONCLUSION Multi agent teamwork is a critical capability in a large number of applications including training, education, ....

Stone, P., Veloso, M.UsingDecision Tree Confidence Factors for Multiagent Control. In Proceedings of the International Conference on Autonomous Agents,1998.

....a wealth of research in game theory and later team theory (e.g. Bascar and Olsder 1982; Greenwald 1998; Owen 1968) but none of this work explicitly considered the temporal effects on the value of information, and how that affected system performance. Recent literature on autonomous agents (e.g. Stone and Veloso 1998; Washington 1998) reports work in distributed, multiagent decision making (using, for example, state space approaches such as Markov Decision Processes) but still fails to consider the temporal issues underlying the usage and value of information. So, while information, as an entity, is used in ....

Stone, P. and Veloso, M., "Using Decision Tree Confidence Factors for Multi-Agent Control," Proceedings of the 2nd International Conference on Autonomous Agents, ACM Press, NY, 1998, pp. 86-91.

....of opponent teams against which to evaluate research. However, the lessons learned by researchers participating in RoboCup, particularly the simulation league, have not often been reported in a form that would be accessible to the research community at large (there are notable exceptions, e.g. [13]) Extracting general lessons in areas of teamwork, agent modelling and multi agent learning is a critical task for several reasons: i) to meet the stated research goals of the RoboCup effort (at least the simulation league) ii) to establish the utility of RoboCup and possibly other common ....

....factors, such as distance, view angle, and view mode (approximating visual focus) All communication between players are done via the server, and are subject to limitations such as bandwidth, range and latencies. Figure 1 shows a snapshot of the soccer server with two competing teams: CMUnited97 [13] versus our ISIS team. Figure 1. The Robocup synthetic soccer domain. In RoboCup97, ISIS97 won the third place prize (out of 32 teams) It won five soccer games in the process, and lost one. In RoboCup98, ISIS98 came in fourth (out of 37 teams) It won or tied seven soccer games in the process, ....

[Article contains additional citation context not shown here]

Stone, P. and M. Veloso: 1998b, `Using Decision Tree Confidence Factors for Multiagent Control'. In: RoboCup-97: The first robot world cup soccer games and conferences. Springer-Verlag, Heidelberg, Germany.

.... competition, agents are to compete through the RoboCup simulator, called Soccer Server [Noda, 1995] Noda et al. 1997] There are many approaches to client implementation, such as Robot Learning Based on LfE Method [Asada et al. 1995] Genetic Programming [Luke et al. 1997] Decision Tree [Stone et al. 1998], and many more. As soccer emphasizes on teamwork, we decided to implement our agent based on multi agent collaboration [Matsubara et al. 1996] Noda et al. 1996] The following sections describe a framework for our agents behaviour, by means of strategical positions, roles and ....

Peter Stone and Manuela Veloso. Using Decision Tree Confidence Factors for Multiagent Control. In RoboCup-97: The First Robot World Cup Soccer Games and Conferences, H. Kitano (ed.), 1998. Springer Verlag, Berlin. Also in Second International Conference on Autonomous Agents, 1998.

....teamwork, agent modeling, and multi agent learning[5] Yet, the lessons learned by researchers participating in RoboCup, particularly the simulation league, have largely not been reported in a form that would be accessible to the research community at large. There are just a few notable exceptions[9]. However, extracting such general lessons in areas of teamwork, agent modeling and multi agent learning is a critical task for several reasons: i) to meet the stated research goals of the RoboCup effort (at least the simulation league) ii) to establish the utility of RoboCup and possibly other ....

....or emergent coordination. Our application learning in ISIS agents is similar to some of the other investigations of learning in RoboCup agents. For instance, Luke et al. 6] use genetic programming to build agents that learn to use their basic individual skills in coordination. Stone and Veloso[9] present a related approach, in which the agents learn a decision tree which enables them to select a receipient for a pass. 7 Lessons Learned from RoboCup Challenges of teamwork and multi agent learning are critical in the design of multi agent systems, and these are two of the critical research ....

P. Stone and M. Veloso. Using decision tree confidence factors for multiagent control. In RoboCup-97: The first robot world cup soccer games and conferences. Springer-Verlag, Heidelberg, Germany, 1998.

....we explore via multiple systems. Our challenge response also attempts to extract general lessons from RoboCup. Indeed, despite the RoboCup aim to stimulate general multi agent research, few RoboCup researchers have extracted domain independent research lessons (there are a few notable exceptions[Stone and Veloso, 1998b] This paper attempts to remedy this situation. 2 Background: Domain and Agents The RoboCup simulation league uses a complex, dynamic, noisy soccer simulation, called the soccerserver, which simulates the players (22) bodies, the ball and the soccer field with goals and flags. Software ....

....6 Lessons Learned Our research in RoboCup has been fueled by the IJCAI 97 challenge. We have responded to the challenge in all three categories of learning, teamwork and agent modeling. Few other RoboCup teams have attacked the IJCAI 97 challenge in this much breadth. One possible exception is [Stone and Veloso, 1998b] who have focused on layered learning for agent design, and an approach to teamwork based on locker room arrangements[Stone and Veloso, 1998a] In their teamwork approach, agents synchronize their individual beliefs periodically in a fixed manner, in contrast with ISIS s STEAM in which ....

[Article contains additional citation context not shown here]

P. Stone and M. Veloso. Using decision tree confidence factors for multiagent control. In RoboCup-97: The first robot worldcupsoccergamesand conferences.Springer- Verlag, Heidelberg, Germany, 1998.

....teamwork, agent modelling, and multi agent learning[6] Yet, the lessons learned by researchers participating in RoboCup, particularly the simulation league, have largely not been reported in a form that would be accessible to the research community at large. There are just a few notable exceptions[11]. However, extracting such general lessons in areas of teamwork, agent modelling and multi agent learning is a critical task for several reasons: i) to meet the stated research goals of the RoboCup effort (at least the simulation league) ii) to establish the utility of RoboCup and possibly ....

....factors, such as distance, view angle, and view mode (approximating visual focus) All communication between players are done via the server, and are subject to limitations such as bandwidth, range and latencies. Figure 1 shows a snapshot of the soccer server with two competing teams: CMUnited97 [11] versus our ISIS team. In RoboCup97, ISIS97 won the third place prize (out of 32 teams) It won five soccer games in the process, and lost one. In RoboCup98, ISIS98 came in fourth (out of 37 teams) It won or tied seven soccer games in the process, and Figure 1: The Robocup synthetic soccer ....

[Article contains additional citation context not shown here]

P. Stone and M. Veloso. Using decision tree confidence factors for multiagent control. In RoboCup-97: The first robot world cup soccer games and conferences. Springer-Verlag, Heidelberg, Germany, 1998.

....is Yokota et al. 20] Our application of learning in ISIS agents is similar to some of the other investigations of learning in RoboCup agents. For instance, Luke et al. 8] use genetic programming to build agents that learn to use their basic individual skills in coordination. Stone and Veloso[16] present a related approach, in which the agents learn a decision tree which enables them to select a recipient for a pass. In terms of related work outside RoboCup, the use of a teamwork model remains a distinguishing aspect of teamwork in ISIS. The STEAM teamwork model used in ISIS, is among ....

P. Stone and M. Veloso. Using decision tree confidence factors for multiagent control. In RoboCup-97: The first robot world cup soccer games and conferences. Springer-Verlag, Heidelberg, Germany, 1998.

....(iii) agent modeling (plan recognition) 15] Yet, the lessons learned by researchers participating in RoboCup, particularly the simulation league, have largely not been reported in a form that would be accessible to the research community at large. There are just a few notable exceptions (e.g. [27]) However, extracting such general lessons in areas of teamwork, agent modeling and multi agent learning, so as to be applicable in other domains beyond RoboCup, is an important task for RoboCup researchers. This article attempts to remedy the above situation by focusing on two of the RoboCup ....

....between two robots. Our application of learning in ISIS agents is similar to some of the other investigations of learning in RoboCup agents. For instance, Luke et al. 17] use genetic programming to build agents that learn to use their basic individual skills in coordination. Stone and Veloso[27] present a related approach, in which the agents learn a decision tree which enables them to select a recipient for a pass. The confidence values from the decision tree are also used to direct the agents actions if no good recipient for a pass is suggested by the decision tree, the agent ....

P. Stone and M. Veloso. Using decision tree confidence factors for multiagent control. In RoboCup-97: The first robot world cup soccer games and conferences. Springer-Verlag, Heidelberg, Germany, 1998.

....Section 3. 2) The formation and each of the units can also specify inter position behavior specifications for the member positions, as illustrated in Figure 6(a) In this case, the formations specify inter role interactions, namely the positions to which a player should consider passing the ball [26]. Figure 6(b) illustrates the units, the roles involved, and their captains. Here, the units contain defenders, midfielders, forwards, left players, center players, and right players. Since the players are all autonomous, in addition to knowing its own role, each one has its own belief of the ....

Peter Stone and Manuela Veloso. Using decision tree confidence factors for multiagent control. In Hiroaki Kitano, editor, RoboCup-97: Robot Soccer World Cup I, pages 99--111. Springer Verlag, Berlin, 1998.

....a set of units. The formation and each of the units can also specify inter position behavior specifications for the member positions, as illustrated in Figure 6(a) In this case, the formations specify inter role interactions, namely the positions to which a player should consider passing the ball [16]. Figure 6(b) illustrates the units, the roles involved, and their captains. Here, the units contain defenders, mid fielders, forwards, left players, center players, and right players. a) Unit = Unit Captain (b) Figure 6: a) A possible formation (4 3 3) for a team of 11 players. Arrows ....

....teammates are said to be passive. We have previously created an action selection algorithm for an active agent in a team that allows for the run time evaluation of what action to take when in possession of the ball: passing the ball to one of the teammates or shooting it directly towards the goal [16]. The interesting question we address here is what should the passive agents do Anticipation allows the passive agents to actually not be passive, but to position themselves with the concrete objective of trying to maximize the chances of a successful pass in case the active agent chooses to ....

Peter Stone and Manuela Veloso. Using decision tree confidence factors for multiagent control. In Proceedings of the Second International Conference on Autonomous Agents, 1998.

....a likely success (ball reaches its destination or a teammate gets it) or likely failure (opponent intercepts the ball) Using the C4.5 DT algorithm [8] the classifications were learned with associated confidence factors. The learned behaviors proved effective both in controlled testing scenarios [9, 11] and against other previously unseen opponents in an international tournament setting [4] These two previously learned behaviors were both trained off line in limited, controlled training situations. They could be trained in such a manner due to the fact that they only involved a few players: ....

....each agent has 8 possible actions as illustrated in Figure 1(a) Since a player may not be able to tell the results of other players actions, or even when they can act, the domain is opaque transition. A team formation is divided into 11 positions (m = 11) as also shown in Figure 1(a) [11]. Thus, the partition function P (s) returns the player s position. Using our layered learning approach, we use the previously trained DT as e. Each possible pass is classified as either a likely success or a likely failure with a confidence factor. Outputs of the DT could be clustered based on ....

Peter Stone and Manuela Veloso. Using decision tree confidence factors for multiagent control. In Hiroaki Kitano, editor, RoboCup-97: Robot Soccer World Cup I, pages 99--111. Springer Verlag, Berlin, 1998.

....there was a player there. Max Range Home Range Center Midfielder, Left Goalie, Home Coordinates Figure 5: Different positions with home coordinates and home and max ranges. which a player should consider passing the ball. We use decision tree learning to help players decide where to pass [36]. Figure 6(b) illustrates the units, the roles involved, and their captains. Here, the units contain defenders, midfielders, forwards, left players, center players, and right players. Unit = Unit Captain (a) b) Figure 6: a) A possible formation (4 3 3) for a team of 11 players. Arrows ....

Peter Stone and Manuela Veloso. Using decision tree confidence factors for multiagent control. In Hiroaki Kitano, editor, RoboCup-97: Robot Soccer World Cup I, pages 99--111. Springer Verlag, Berlin, 1998.

....research in the robotic soccer domain within the RoboCup initiative [5] which, in 1997, included a simulator league and small size and medium size robot leagues. We have been doing research extensively in the simulator league, developing learning techniques and team strategies in simulation [12, 14]. Many of these team strategies were directly incorporated into the robotic system described here. We are currently also applying machine learning techniques to acquire hard to tune boundary behaviors for the real robots. This paper describes the overall architecture of our robotic soccer team. ....

Peter Stone and Manuela Veloso. Using decision tree confidence factors for multiagent control. In Hiroaki Kitano, editor, RoboCup-97: The First Robot World Cup Soccer Games and Conferences. Springer Verlag, Berlin, 1998. In Press.

....research in the robotic soccer domain within the RoboCup initiative [7] which, in 1997, included a simulator league and small size and medium size robot leagues. We have been doing research extensively in the simulator league, developing learning techniques and team strategies in simulation [14, 16]. Many of these team strategies were directly incorporated into the robotic system described here. We are currently also applying machine learning techniques to acquire hard to tune boundary behaviors for the real robots. In this paper, we focus on presenting our team of small robotic agents which ....

Peter Stone and Manuela Veloso. Using decision tree confidence factors for multiagent control. In Hiroaki Kitano, editor, RoboCup-97: The First Robot World Cup Soccer Games and Conferences. Springer Verlag, Berlin, 1998. In Press.

....research in the robotic soccer domain within the RoboCup initiative [6] which, in 1997, included a simulator league and smallsize and medium size robot leagues. We have been doing research extensively in the simulator league, developing learning techniques and team strategies in simulation [12, 11]. Many of these team strategies were directly incorporated into the robotic system described here. We eventually hope also to transfer these learning techniques to the real system as we develop a complete Robotic Soccer architecture. In this paper, we focus on presenting our team of small robotic ....

Peter Stone and Manuela Veloso. Using decision tree confidence factors for multiagent control. In Proceedings of the First International Workshop on RoboCup, Nagoya,Japan, August 1997.

....or a teammate gets it) or likely failure (opponent intercepts the ball) Using the C4.5 DT algorithm [Quinlan, 1993] the classifications were learned with associated confidence factors. The learned behaviors proved effective both in controlled testing scenarios [Stone Veloso, 1998a, Stone Veloso, 1998c] and against other previously unseen opponents in an international tournament setting [Kitano et al. 1997] These two previously learned behaviors were both trained off line in limited, controlled training situations. They could be trained in such a manner due to the fact that they only ....

....each agent has 8 possible actions as illustrated in Figure 1(a) Since a player may not be able to tell the results of other players actions, or even when they can act, the domain is opaque transition. A team formation is divided into 11 positions (m = 11) as also shown in Figure 1(a) Stone Veloso, 1998c] Thus, the partition function P (s) returns the player s position. Using our layered learning approach, we use the previously trained DT as e. Each possible pass is classified as either a likely success or a likely failure with a confidence factor. Outputs of the DT could be clustered based on ....

Stone, P., and Veloso, M. 1998c. Using decision tree confidence factors for multiagent control. In Kitano, H., ed., RoboCup-97: Robot Soccer World Cup I. Berlin: Springer Verlag. 99--111.

....Section 3. 2) The formation and each of the units can also specify inter position behavior specifications for the member positions, as illustrated in Figure 5(a) In this case, the formations specify inter role interactions, namely the positions to which a player should consider passing the ball [18]. Figure 5(b) illustrate the the players would pass to the fixed positions regardless of whether there was a player there. Max Range Home Range Home Coordinates Center Midfielder, Left Goalie, Figure 4: Different positions with home coordinates and home and max ranges. units, the roles ....

P. Stone and M. Veloso. Using decision tree confidence factors for multiagent control. In H. Kitano, ed, RoboCup-97: The First Robot World Cup Soccer Games and Conferences. Springer Verlag, Berlin, 1998, forthcoming.

No context found.

P. Stone and M. Veloso. Using decision tree confidence factors for multi-agent control. Autonomous Agents, 1998.

No context found.

Peter Stone and Manuela Veloso, `Using decision tree confidence factors for multiagent control', in Second International Conference on Autonomous Agents, (1998).

No context found.

Peter Stone and Manuela Veloso. Using Decision Tree Confidence Factors for Multiagent Control. In RoboCup-97: The First Robot World Cup Soccer Games and Conferences, H. Kitano (ed.), 1998. Springer Verlag, Berlin. Also in Second International Conference on Autonomous Agents, 1998.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC