Before we present the oordination model, we will introdu e the Pyrosim Agent platform [7℄ that we is limbing a hill it annot see to the other side of the hill). Agents also re eive information about state (physi al energy, speed, a eleration, position in the terrain, status of the personal water jet) as well. Additional information about the Pyrosim platform and simulation model may be found a visualization of a simulation in the Pyrosim simulator. Pyrosim reates a omplex environment at [7℄. as well as several matrix stru tures named Visual Maps that des ribe lose range and medium range a team of Agents (regh ters) ooperates to ontrol and extinguish the re, while simultaneously need for ooperation. Agents may ommuni ate with ea h other in order to organize team eorts that may be used as testbed for team oordination models, and for simulating regh ting ta ti s olleagues to remain safe. Agents are equipped with a water jet with limited power that allows have been using in our experiments. The Pyrosim platform simulates a forest-re environment where them to put out the re, but they are not normally able to do it individually so there is an obvious trying to minimize the overall damage and losses. In Pyrosim, Agents have to deal with very dynami re fronts, terrain onstraints and their own physi al and logisti limitations. Ea h Agent ells where no information is available be ause of the o lusion ee t (for instan e, when an agent needs to ensure its physi al safeness while trying to gh t the re and, at the same time, help visible parameters of other Agents (lo ation, approximate energy level and a tion). Figure 1 shows on the distan e) about terrain, vegetation, level of destru tion, and re ells. Visual Maps may have surroundings. Visual Maps ontain information with dieren t levels of detail and noise (depending (broad ast and 1-to-1 messages). Agent’s Per eption System provides information about his own global knowledge. In a de entralized solution, the information would be dispersed, and no ommuni ations between agents. every agent to be aware of the other agents’ knowledge, but that would in rease drasti ally agent would have a omplete understanding of the situation. The alternative would be for One agent, the Leader, may have a ess to the global situation and make de isions based on • will instantiate it to the forest regh ting domain.
In the next subse tions we will make a formal des ription of the proposed model and then we the Leader agent’s "personality in order to obtain a more autious behavior. As result the Leader Leader fails, the team oordination is lost. To prevent the team from losing its Leader we ongured is no me hanism in our implementation for repla ing the Leader in ase of Leader failure. However, this is not a big issue be ause our goal is to simulate regh ting ta ti s and not to reate a system for using during real regh ting. of agents. Another problem with entralized oordination is that if ommuni ation fails or if the keeps a higher distan e from the ames and takes less han es during the regh t. Currently, there dieren t ongurations (re size, shape and properties may vary a lot). is not xed (on the other hand, so er teams have xed size), and the opponent may have very prin iple of these teams, we had to build a model that would be able to handle some issues that are Roles is not related with the number of agents. Teams may be homogeneous (all agents have the other hand, the so er eld has xed dimensions), the number of agents that onstitutes a team We will now present a formal des ription of the proposed model. Every agent has a Role (3).
There is a predened number of Roles (2) whi h dene dieren t agent behaviors. The number of so er teams by Stone and Veloso [8, 9℄, and by Reis and Lau [4, 5, 6℄. Both approa hes dened the generi team formations that may be used with any team of agents. Despite we are using the same team spatial distribution using roles, whi h were then assigned to agents. Roles enable to dene For developing the oordination model, we used the same prin iple that was used in the roboti not present in the roboti so er domain, su h as the a tuation area size is not delimited (on the In the proposed model, an Atta k Plan denes a sequen e of Tasks for a given Role (7). In this way, agents that share the same role will have the same tasks to arry out. In general, an Atta k Plan may be arried out by one Role only, although there are Atta k Plans that may be exe uted by a given set of Roles (6). ta ti s Dynami Ta ti s (10), and they are omposed by a set of Ta ti s with A tivation Conditions distribution using absolute quantities (e.g. allo ate 2 agents to a given Role). Additionally, agent half of the team to a given Role). However, in our implementation, it is also possible to dene agent assignments. (11). A tivation Conditions dene when to swit h to a given ta ti . distribution may be dened dynami ally at run-time, whi h enables to produ t more omplex role It is also possible to dene Ta ti s that hange the team approa h over time. We all these From (9), we gather that agent assignment to Roles is dened using relative quantities (e.g. allo ate AttackP lans = {AttackP lan1, . . . , AttackP lansnattacks}
T asks = {T ask1, T ask2, T ask3, . . . , T askntasks}
AdmissibleRolesi ⊆ Roles ∀i = 1..nattacks
AttackP lani = {Rolej, AttackT ask1, . . . , AttackT asknattacktasks} ∀i = 1..nattacks Rolej ∈ AdmissibleRolesi AttackT askk ∈ T asks
T actics = {T actic1, T actic2, T actic3, . . . , T acticntactics} DynamicT actics = {DynamicT actic1, . . . , DynamicT acticndynamictactics} DynamicT actici = {ActivationCondition1, SubT actic1, ...,
ActivationConditionnsubtactics, SubT acticnsubtactics} ∀i = 1..ndynamictactics SubT acticj ∈ T actics swat the ames. Therefore, this kind of Atta k Plans spe ify tasks for approa hing the re and then have divided Atta k Plans in two major lasses: (i) Dire t Atta ks, and (ii) Indire t Atta ks. This when the re is too intense for regh ters to approa h it. The most ommon te hnique is to build Atta ks involve gh ting the re dire tly in the ames using water or manual tools, like shovels, to represent abstra t lasses that are used for stru turing proposes only. In general terms, Dire t Atta ks dene how to atta k a given part of the re (e.g. the head, the tail or the anks) and the relines. However, we implemented another kind of indire t atta k that onsists in reating a wetwith no oloring represent Atta k Plans that may be instantiated, and the lasses olored in grey WetLineAtta k denes how to reate a Wet-Line in a given area. supports dogging operations. ee t is similar to building a reline, although this is just a temporary solution until the simulator For instantiating the proposed model to forest regh ting, we start by dening Atta k Plans. We Figure 2 illustrates a lass diagram of the Atta k Plans that we implemented, where the lasses atta k it dire tly. On the other hand, Indire t Atta ks are used to gh t re at distan e, espe ially division is based on regh ting theory that dieren tiates Dire t from Indire t Atta ks [3℄. Dire t line. Instead of dogging the ground, regh ters wet the ground with large quantities of water. The re to starve it out of fuel. In the urrent version of the re simulator it is not possible to built a r eline. Firelines are built by dogging the ground to remove all the vegetation in front of the s enario B regh ters arrive to the re at a later stage where using a dire t atta k is not enough a hieve negative results in situations where others su eed [2℄. We are able to observe this in the regh ters start by atta king re dire tly and therefore they are able to prevent it from spreading at in order to reate a Wet-Line. We have sele ted these ta ti s in order to represent (i) ta ti s (gure 5). ST1 ta ti was a good hoi e for s enario A but that is not the ase for s enario B. In sa ri ing the same area in both s enarios. In this ta ti regh ters build a wet-line around the re area that prevents re from spreading, instead of gh ting the re dire tly. Therefore, the total in s enario B, and therefore when regh ters arrived to the re in s enario B they would fa e a re to ontrol re. On the other hand, ta ti ST6 was able to a hieve a more stable performan e by Currently, the developed system works as a platform to test ta ti s in dieren t s enarios. However, ST6) in two dieren t s enarios (A and B). Ta ti ST1 tries to atta k the re by positioning all appli ability of dire t and indire t atta ks. we know from regh ting theory that should be applied to res with dieren t dimensions (among an earlier stage. However, in s enario B this same ta ti (ST1) failed to ontrol re from spreading (gures 3 and 4) and that ta ti ST1 a hieved better results. This happened be ause in ta ti ST1 We experimented both ta ti s in two s enarios with the same hara teristi s (in terms of terrain More information about this and other experiments may be found in [2℄. An important output geometry, vegetation and weather), but in s enario A agents re eived the warning sign sooner than our next step is to enhan e the platform in order to automati ally determine the ta ti s that Ta ti ST6 pla es regh ters around the re perimeter and give them orders to wet the ground from these experiments is that results are oherent with forest regh ting theory in terms of the following example where we present the results of experimenting two dieren t ta ti s (ST1 and other fa tors su h as the number of regh ters available, terrain properties and weather onditions). in a more advan ed state than in s enario A. We observed that both ta ti s su eeded in s enario A burned area in both s enarios was nearly the same, but in s enario A the damage was ex essive. mainly based in dire t atta ks (ST1), and (ii) ta ti s mainly based in indire t atta ks (ST6), whi h work best in dieren t s enarios. We have already made some experiments that show that there agents behind the tail of the re and giving them orders to atta k the re dire tly with water. are ta ti s that a hieve positive results in situations where others fail, although the same ta ti s 169 173 186 Time Elapsed 2.5
Fire not attacked
ST1 with 4 agents 2 SSTT11 wwiitthh 86 aaggeennttss ) s ll (ce 1.5 a e r A edn 1 r u B 0.5 0 0 606.30 Burned Area 625.49 569.60 In this paper we have presented a model for oordinating a team of agents. We have instantiated this model to the forest regh ting domain where we were able to implement some ta ti s similar ta ti s in a rather exible way. In our implementation, the overall team oordination is ontrolled to the ones that real regh ting teams use. The model we presented enables to dene mu h other 273, June 1999. bandwidth ommuni ation for real-time strategi teamwork. Arti ial Intelligen e, 110(2):241 [9℄ Peter Stone and Manuela Veloso. Task de omposition, dynami role assignment and low[10℄ Mar o Wiering and Mar o Dorigo. Learning to ontrol forest res. In H. Haasis and K. Ranze, mental Prote tion’, pages 378388, 1998. editors, Pro eedings of the 12th international Symposium on ’Computer S ien e for EnvironCarnegie Mellon University, De ember 1998. [8℄ Peter Stone. Layered Learning in Multi-Agent System. PhD thesis, S hool of Computer S ien e, and more experiments should be done. to regh ting theory. However, we are aware that we are still far from providing meaningful We also pointed out an interesting line of resear h regarding possible uses of ma hine learning ta ti s that performed best in the tested s enarios are the ones that we were expe ting a ording We also presented some results of our experiments that demonstrate that, like in reality, dieren t for automati ta ti sele tion based on the results of ta ti experimentation in dieren t s enarios. by a single agent, the Leader, who is responsible for arrying out ta ti s. However, we have shown information for real regh ting teams. For a hieving this, the simulator should be properly validated re s enarios require dieren t regh ting ta ti s in order to minimize re damage. Additionally, the that agents have lo al autonomy and are able to ooperate lo ally without the Leader intervention.
RoboCup to Real-World Appli ations (sele ted papers from the ECAI 2000 Workshop and additional ontributions), pages 175197, London, UK, 2001. Springer-Verlag.
Pagello, editors, Balan ing Rea tivity and So ial Deliberation in Multi-Agent Systems, From oordinating a team of homogeneous agents. In Markus Hannebauer, Jan Wendler, and Enri o [6℄ Lus Paulo Reis, Nuno Lau, and Eugenio Oliveira. Situation based strategi positioning for [4℄ Lus Paulo Reis. Coordenaªo em Sistemas Multi-Agente: Apli aı es na Gestªo UniversitÆria e Futebol Robti o. PhD thesis, Fa uldade de Engenharia da Universidade do Porto, Porto, Portugal, June 2003. the Pyrosim Agent Platform and of the Firegh ting Agents Ar hite ture.
The authors would like to thank to Lus Sarmento for his great ontribution in the development of [3℄ Mar Ni olas and Grant Beebe. The training of forest regh ters in indonesia. Te hni al report, German Agen y for Te hni al Cooperation, European Union, and Government of Indonesia, 1999. http://d m.web.simplesnet.pt/publi ations/MS _d m_0 6.pdf. main. Master’s thesis, Fa uldade de Engenharia da Universidade do Porto, 2006. [2℄ Daniel Moura. Coordinating a team of agents in the forest regh ting dohampion. In Peter Stone, Tu ker Bal h, and Gerhard Kraetzs hmar, editors, RoboCup 2000: [5℄ Lus Paulo Reis and Nuno Lau. F portugal team des ription: Robo up 2000 simulation league Robot So er World Cup IV, pages 2940, London, UK, 2001. Springer-Verlag. [7℄ Lus Sarmento. An emotion-based agent ar hite ture. Master’s thesis, Fa uldade de CiŒn ias da Universidade do Porto, 2004.
Te hni al report, MinistØrio da Agri ultura do Desenvolvimento Rural e das Pes as, Portugal, January 2006. http://www.dgrf.min-agri ultura.pt/v4/dgf/pub.php?ndx=2271. [1℄ Divisªo da Defesa da Floresta Contra In Œndios. In Œndios orestais relatrio de 2005.
A knowledgements Referen es