The Third International Conference on Cloud and Robotics ( ICCR 2016 ) , Sain t Quentin, France, 22 - 23 November 2016  Context Reasoning in Underwater Robots Using  MEBN  Xin Li,   José - Fernán Martínez , Gregorio Rubio   and David Góme z  Centro de Investigación en Tecnologías Software y Sistemas Multimedia para la Sostenibilidad (CITSEM)  Universidad   Politécnica de Madrid  Ctra. de Valencia, Km. 7 28031   –   Madrid ( Spain )  E - mail: xin.li@upm.es, jf.martinez@upm.es, gregorio.rubio@upm.es ,   david.gomezs@upm.es  Abstract — This   paper   presents   ongoing   research   in   the  SWARMs   project   towards   facilitating   context   awareness   in  underwater robots.   In particular, the focus   of   this paper   is put on  the   context   reasoning   part.   The   underwater   environment  introduces uncertain ties in context data which lead   to difficulties  in   the context reasoning phase .   As probability is the   best   well -  known formalism   for computational scient ific reasoning under  uncertaint ies ,   t he emerging   and effective   probabilistic reasoning  method,   namely,   Multi - Entity   Bayesian   Network (MEBN) ,   is  explored for its feasibility   to reason   under uncertainties   in the  SWARMs project . A simple use case   for   oil spill monitoring is  used to v erify   the usefulness of MEBN. The results show that the  MEBN is a promising approach to reason about   context   in the  presence of uncertainties in the underwater   robot field.  Keywords — context   awareness ;   underwater   robots ;   MEBN ;  uncertainty ;   reasoning  I.   I NTRODUCTION  The work presented in this paper is part of the SWARMs  (Smart and Networking Underwater Robots in Cooperation  Meshes) project   ( http://www.swarms.eu/ ) ,   a European   project  in the area of underwater robotics.   The   aim of   the   SWARMs  project   is   to expand the use of underwater robots (e.g., AUVs,  R O Vs ) in maritime and offshore operations and enable them  to work in a cooperative , coordinative,   and context - aware  manner.   The SWARMs approach is underpinned by designing  a semantic and distributed middleware. The middleware layer  is   basically designed   to   facilitate   communication between  heterogeneous   vehicles   and   Mission   Management   Tool  (MMT) which is located in   Comma nd   &   Control Station (C   &  CS) and provide a set of common services. The middleware  features a context - aware framework which is dedicated to  deliver ing   context awareness in underwater robots.  Context awareness is a significant feature that could enable  und erwater robots to understand the complete picture of the  underwater   environment   and   adapt   their   behaviors  accordingly.   Context   awareness   implies   an   effective  exploitation of context.   To   maximize the use of context ,   t he  context - aware framework is envisaged   to   provide well - defined  context treatments , such as abstracting context heterogeneity,  enabling   sharing and   re use of context   &   capabilities between  AUVs/ROVs,   delivering   useful   information   to   robots   &  operators for decision making etc.  The services provide d by the context - aware framework  generally comply with the lifecycle of context awareness.  According   to   Perera   et   al.   [1] ,   t he   lifecycle   of   context  awareness mainly consists of four phases, including context  acquisition, context   mode ling , context reasoning, and context  dissemination.   Especially, context reasoning is   a   significant  process   becaus e   underwater   robots or   operators are   more  willing to use high - level context for decision - making.   For  instance, a piece of high - level context   information ,   vehicle A  might collide with vehicle B soon ,   deduced from   available  context   information   vehicle A is o ut of trajectory and heading  in the direction of vehicle B , could enable operators with a  better understanding of the situation .   In the SWARMs project,  a variety of context, referring to any information that could be  used to character ize underwater robot s and their operational  environment,   could   be   obtained,   such as battery level   of  vehicles, capabilities of vehicles, currents, turbidity, salinity,  and   wind   direction   etc.   Considering   the   harsh   nature   of  underwater environments, context   is   prone to be unce rtain   due  to   unreliable   communications,   data   loss,   imperfect  instruments, or partial views, etc   [2] .   The uncertainty, as an  inherent characteristic of context, introduces more challenges  in   context reasoning in the   underwater robot field .  Ontological reasoning and rule - based reasoning are widely  used   context   reasoning   methods.   However,   they   lack   the  capability of reasoning under uncertainties as they assume a  deterministic   world.   Probabil ity   is   the   best   well - known  formalism   for   computational   scientific   reasoning   under  uncertainties.   As   a   widely   used   probabilistic   reasoning  method, Bayesian networks (BN) could represent and reason  about problems involving many related hypotheses. However,  i n a standard BN, all the hypotheses and dependencies must be  fixed and only the evidence could be varied from problem to  problem. However,   many real - world   problems   could involve   a  varying number of entities. Thus, standard BN cannot be able  to handle uncer tainties in such real problems.   M ulti - E ntity  B ayesian   N etwork (MEBN)   [3]   extends   BN to achieve first -  order expressive power and it provides a compact way to  represent repeated structures in a BN. MEBN   has shown its  usefulness   in   reasoning   under   uncertain ties   in   many  applications,   such   as   procurement   fraud   detection   [4] ,  maritime   awareness   [5] ,   knowledge - driven   analysis   for  cultural heritage   [6] .   In the SWARMs project, there may exist  a lot of   complex   reasoni ng problems which include   a   varying  number of entities. For example,   the   investigated   robot   may  have more than one robot around. The collision risk is affected  by all possible nearby robots, thus, resulting in a problem  which involves   a   varying number of entities .   The MEBN
The Third International Conference on Cloud and Robotics ( ICCR 2016 ) , Sain t Quentin, France, 22 - 23 November 2016  Fig. 1.   The   proposed context - aware framework .  shows   the   potential   in reasoning   about   such complex problems  in the SWARMs project.   Therefore,   it is worth exploring the  usefulness of the MEBN   in the SWARMs project .   This   paper  presents   the   conceptual proposal of   context - aware framework  defined in the SWARMs projec t   and   explor es   the applicability  of   the MEBN   in context reasoning   using a case study . The  usefulness of   MEBN is verified by r easoning the severity level  of   a spilled region   in an oil spill monitoring use case.  II.   T HE   P ROPOSED   C ONTEXT - A WARE   F RAMEWORK  The cont ext - aware framework, as part of the SWARMs  middleware,   is shown   in fig .   1.   It   is conceived to be modular  and distributed which could imply a better potential over a  centralized one to be employed in robots coordination and  cooperation.  T he   context - aware   framework   consists   of   six   logic  components,   namely,   Data   Processor ,   Semantic   Mapper ,  Ontology   Model ,   Rules   Creator ,   Context   Reasoner ,   and  Semantic Query . Functional capabilities of each component  are descri bed in the following sections.  Da ta   Processor .   T he   extreme   underwater   conditions  impose more challenges in the data acquisition phase   and  c ontext   data   obtained   from   the   environment   could   be  uncertain. A preliminary   data   treatment is provided by the data  processor component. This component can pre - process data  received from   sensors   or other sensing instruments. Statistics  on historical data, such as probability distribution, can be  learned using machine learning algorithms   in this component.  Operations,   such   as   validating   values,   checking  inconsistencies,   calculating   uncertainty   degrees,   removing  outliers or filling in missing values etc., can be executed in  this component.  Ontology Model . This component, acting as a semantic  repository, keeps the SWARMs ontology and stores all data  obtained as instances in the ontology model. Ontology is  adopted to serve as a common information model to represent  information and enable sha ring, reuse, and integration of data  between vehicles. The information model is structured in a  hierarchical   manner.   The   networked   information   model  consists of three levels of ontologies: core ontology (acting as  an   upper - level   ontology   to   glue   all   domain   specific  ontologies), domain specific ontology (providing information  models for different domains, e.g., mission planning,   robotic  vehicles,   environment   recognition   &   sensing ,   and  communications   &   networking ),   and   application - specific  ontology (describin g information with a focus on particular  applications). Therefore, data, including, but not limited to  contextual   measurements,   vehicle - related   data,   data   from  external sources (oceanic weather forecast etc.), and marine  experts knowledge, can be abstracte d and formalized in an  ontological   format.   All   data   can   be   displayed   with   a  homogeneous view associated with the semantic content. The  ontology model   could   annotate   probability information   of  uncertain context by using ontology constructs defined in the  PR - OWL   [3]   ontology.   With   the   PR - OWL   classes   and  properties, ontology engineers could fully specify   a n   MEBN  model   using   the   OWL   ontology   language.   Thus,   the  SWARMs   ontology   could   not   only   annotate   probability  information about uncertain   context   but also provide support  for the MEBN reasoning.  Semantic Mapper . This component plays a vital role in  facilitating the transparent sharing of information. As data  might be formatted in different manners pertaining to different  data sources, this componen t aims to parse them and formalize  them in an ontology compliant   format. Translations from  different standards, such as XML, JSON, or binary files to  ontological formations (RDF, or OWL) can be enabled in this  component.   For   differently   formatted   data   whic h   do   not  comply with the SWARMs ontology, corresponding mapping  files should be predefined in order to parse and map them into  the common information model.  Rules Creator . Operators or marine experts are able to  define rules according to their knowledge an d changes in the  environment through this component. A set of user - defined  rules can be   translated into the SWRL (Semantic Web Rule  Language) format and   inserted into the   ontology model . Rules  can be diverse, including restrictions or definitions for entit ies,  regulations for evaluating data values, or specifications for  relationships between entities. The rule set is very important   to  be considered by the   context reasoner to make inferences.  Context   Reasoner .   The   essential   capability   of   this  component is t o   deduce   new knowledge   based on available  context stored in the   ontology model . Basically, it will consult  information   stored   in   the   ontology   model   and   also   take  experience s   and knowledge from marine experts into account.  A hybrid reasoning mechanism , incl uding ontological, rule -  based and MEBN reasoning,   is   intended to be   adopted by this  context reasoner.   Specifically, o ntological reasoning enables  several kinds of operations, including concept satisfiability,  consistency check, class subsumption and logic   inference.   The  r ule - based   reasoning could augment ontological reasoning in  terms   of   logicality   and   human - readability.   The   MEBN   is  dedicated to reasoning under uncertainties and it is the focus  of this paper which will be elaborated in   section   III   and   IV .  Semantic Query . This component deals with any semantic  query   made   by   operators.   It   receives   queries,   calls  corresponding reasoning services,   and finally   output s answers.  This is one of context dissemination strategies defined in the
The Third International Conference on Cloud and Robotics ( ICCR 2016 ) , Sain t Quentin, France, 22 - 23 November 2016  Fig.   2 .   Classical BNs to reason about the scenario .  SWARMs   project . Apart from this, the SWARMs   project   also  follows   a   publish ing / subscription   paradigm   to   distribute  context information.  The   context - aware   framework   provides   architect ural  support for different context treatments. In particular, t he  context reasoner is   of significance as i t is expected to   provide  necessary services to meet   all the   reasoning   requirements in  the   SWARMs   project .   Especially,   it   should   handle   the  intricacies   adherent to the reasoning over uncertainties .   The  feasibility of integrating MEBN in the context reasoner is  explored in the following sections.  III.   P RINCIPLES OF   MEBN  MEBN unifies Bayesian probability and statistics with  classical   first - order   logic.   By   incor porating   Bayesian  probability with first - order logic, the MEBN   is able to provide  a consistent treatment   of   uncertainty , such as representing  uncertainty about the type of an entity, refining type - specific  probability   distributions   through   Bayesian   Learnin g,   and  reasoning   under   uncertainties .   Beyond   the   capability   of  traditional   BN   in   reasoning   about   the   fixed   number   of  attributes, MEBN could deal with   a   varying number of entities  with   their   number,   type,   and   relationships   undetermined.  MEBN provides a mean s of defining probability distributions  over   an   unbounded   and   varying   number   of   interrelated  hypotheses with the aid of syntax, a set of model construction  and inference processes, and semantics.  MEBN interprets the world as a set of entities t hat have  attributes and have ca u s al relationships with other entities.  Knowledge   about   t he   attributes   of   the   entities   an d   their  relationships to each other is represented as   a n   MEBN   model.  MEBN logic consists of a collection of MEBN fragments  (MFrag) organized into   an   MEBN Theory (MTheory).   An  MTheory could represent a particular domain of discourse.  Each MFrag, as a modular component, represents knowledge  about specific subjects within the domain of discourse and it  models   probability   information   about   a   group   of   r elated  random variables.   Similar to BN, each MFrag is a Directed  Acyclic Graph (DAG) with parameterized nodes that represent  attributes of entity and edges that represent dependencies  amongst them. Three types of nodes, including   resident nodes ,  context no des ,   input nodes , are defined in the MTheory.     Resident nodes.   In an MFrag graph, resident nodes  represent   variables   that   have   local   probability  distributions dependent on the values of their parents.  Exactly one home MFrag is assigned to contain the  comple te expression of a resident node.     Context   nodes.   Context   nodes   are   Boolean   nodes,  including value True, False, and Absurd. The MFrag  must satisfy the conditions expressed by context nodes  in order to be valid.     Input   nodes.   Input   nodes   have   their   distributi ons  defined in other MFrags and they are important inputs  for the definition of resident nodes.  The   MTheory   could   be   instantiated   with   specific  information about the individual entity instances to reason  about   specific   situations.   The   instantiate d   MEBN,   na med  Situation   Specific Bayesian Network (SSBN), could make the  inference under uncertainties based on standard BN reasoning.  The   open source UnBBayes   [7]   tool   is available to provide  both   GUI and Java API   for   build ing   MEBN and execut ing  learning and reason ing.  IV.   U SING   MEBN   IN AN   O IL   S PILL   M ONITORING   P ROBLEM  In this section, a   case study on   oil spill monitoring   is  presented. Through this case study, the comparison between  the BN and the MEBN reasoning is provided and also the  applicability of the MEBN reasoning is verified.  A.   Description of the   O il   S pill   M onitoring   S cenario  O il spill detection is considered as one   of use cases in the  SWARMs   project.   Proactive   detection   of   oil   spills   is   an  important means to minimize the damage of spills to the  marine   environment.   A   set   of   SWARMs   vehicles   could  collaboratively detect the occurrence of oil spills based on  contextual   data they could obtain. After detecting oil spills in a  specific marine region, it is significant to predict the severity  level of the spilled region so that   op erators could have a better  understanding   of   the situation and make mission planning/re -  planning   (e.g., clean - up, containment)   for underwater robots.  The estimation of the severity level of the specific area is  subject t o several available   context s , such as   the   thickness   of  spills,   the   estimated   size of spills, weather conditions, and  currents. In ad dition, the severity level of the specific area  could be influenced by different spills concurrently occurring  in the same area.  B.   BN for   R epresenting   the Scenario  Fig. 2(a) is the BN that can be used to represent the severity  level and its influencing variables. It expresses that currents,  weather, thickness, and estimated size are the main criteria to  determine the severity level of a spilled region. Fig. 2(a) show s  the major shortage   of   the BN is that it cannot serve if there is  any change to the scenario.   For instance, if   two oil spills  (spill_1   and   spill_2)   are   detected   in   a   specific   sea   area  (region_1). Spill_1 is observed as being thick and large and  spill_2 is   detected as being thin and small. In addition, the  weather condition above the spilled marine region is known to  be inclement. Currents in the region_1 fluctuate very strongly.  The BN ,   shown   in fig. 2(a) ,   cannot acc ommodate to this  situ ation which contains   two spill instances.   It is necessary to  define   a   new BN,   illustrated   in   fig.   2(b) ,   to   provide   the
The Third International Conference on Cloud and Robotics ( ICCR 2016 ) , Sain t Quentin, France, 22 - 23 November 2016  Fig.   6 . T he SSBN for the scenario . Fig.   4 .   The local probability distribution of SpreadSpeed .  Fig.   3 .   The   proposed MEBN model .  corresponding representation and reasoning.   When changes  occur in the scenario (e.g., three, or more   spills   exist in the  same region) , BN models need to be revised.   T he BN ,   shown  in fig. 2(b) ,   repeatedly models elements from the BN shown in  fig. 2(a) with the same semantics.   From a domain   mode ling  perspective, a more generic and flexible approach is needed to  allow instantiation of nodes   in order to adapt to different  situation.  C.   MEBN   for   Representing the S cenario  A n   MEBN   model , shown in fig .   3 ,   is built using the  UnBBayes tool   to represent the described reasoning problem  in a modular way .  T he   MTheory   consists   of   seven   MFrags,   including  Thickness ,   Estim atedSize ,   Weather ,   Currents ,   Location ,  SpreadSpeed ,   and   SeverityLevel .   Each   MFrag   represents  knowledge   for   a   specific   entity   and   its   local   probability  distribution.   For instance, the   Sp r eadSpeed   MFrag models that  the spread speed of detected spills is affected by   two inputs,  namely,   we ather and currents. These two input nodes,   weather  and   currents,   are   specified   in   the   Weather   and   Currents  MFrags, respectively.   The local probability distribution of  SpreadSpeed   is   shown   in   fig .   4 .   For   the   demonstration  purpose,   all   the   probability   information   in   this   model   is  directly   provided   by   marine   experts   based   on   their  experiences.  With the seven MFrags, the MTheory could collectively  model   the   unique   joint   probability   distribution   for   the  SeverityLevel   entity.   This   MTheory could be   instantiated   to  specific   scenario s and make inferences   based on observed  evidence   correspondingly .   For   instance,   given   the   same  findings   (shown in fig. 5)   which are   described in section B, the  MEBN   could be specialized to the described scenario, namely,  SSBN, and it is shown in fig.   6 .  With the   evidence   listed in fig. 5,   the SSBN can answer  specific questions of interest . For instance, b ased on standard  BN reason ing, t he SSBN   could infer   that   the region_1 with   the  existence   of   two spills   could be estimated as very serious with  a probability of 88%.   This high - level context could be   sent to  the MMT   for further exploitation . For instance, it could be used  by operators as a parameter to conceive a plan   for underwater  robots so that they can   take remedial measures accordingl y. In  this   situation , only two   spill   instances   are   exemplified as  influence factors for the severity level. It is worth noting that  the built MEBN could be flexible to accommodate to reasoning  which   involves   more   spill   ins tances.   Therefore,   the   same  MEBN   could   be   instantiated   when   dealing   with   varying  number of entities.  Fig.   5 . Findings and query in the scenario.
The Third International Conference on Cloud and Robotics ( ICCR 2016 ) , Sain t Quentin, France, 22 - 23 November 2016  D.   Discussion  Through th is   case study, the MEBN reasoning has shown  its flexibility in adapting different situations over standard BN  reasoning. It provides a   compact way to represent repeated  structures in a BN. Compared with BN, there is no fixed limit  on the number of entity instances in the MEBN model, thus, it  could allow dynamic instantiation of variables to accommodate  the   different   situation.   Another   ad vantage   of   the   MEBN  reasoning is that   MEBN models   could be specified in an  ontological language   (e.g., OWL)   using ontology constructs  provided by the PR - OWL ontology. This feature is particularly  attracting to the SWARMs project as ontology is the chosen  s olution to represent heterogeneous context in the cooperation  of underwater vehicles.   Thus, the MEBN reasoning could  easily be applied to the SWARMs ontology and deduce more  useful information. However, similar to the BN reasoning , the  construction of MEBN   models is highly dependent on the  knowledge of domain experts. It implies huge demands for  exhaustive and exclusive hypotheses.   Also different domain  expert   might   have   a   different   interpretation   for   the   same  reasoning problem, thus, leading to   the   differe nt   mode ling   of  MEBN for the same reasoning problem.   In addition, to ensure a  good   approximation   to   the   real   probability   distribution,   a  massive amount of historical data is needed.  V.   C ONCLUSIONS  This paper has presented the curr ent research   conducted   in  the SWARMs project on facilitating context awareness in  underwater robots.   A conceptual proposal   for   context - aware  framework has been presented. The proposed context - aware  framework aims to provide a complete context management  which offers different da ta treatments complying with the  general lifecycle of context awareness.   In the underwater robot  field,   c ontext   reasoning,   as   a   significant   phase   to   realize  context   awareness,   is   more   difficult because   context   data  obtained by robots are prone to be uncert ain. To tackle the  challenge   in   reasoning   under   uncertainties,   the   emerging  probabilistic reasoning method, MEBN, has been studied and  its applicability in the SWARMs project has been explored.  The MEBN   has been used to reason about the severity level of  a   spilled marine region in an oil spill monitoring scenario. The  results have shown that the MEBN reasoning could be feasible  to   make   inferences   under   uncertainties   in   such   complex  reasoning problems in underwater robots.  Future work would be focused on th e following aspects:     The presented context - aware framework is still at the  prototypical level. To complete its implementation  and testify it in real scenarios will be the next step  forward.     The MEBN reasoning   will be tested with more use  cases and the   integration   with   the   ontological   and rule -  based reasoning in the context reasoner   is intended as  future work .     Context processing and reasoning are normally time  and resource consuming. How to achieve a trade - off  between the load on managing context data an d the  actual benefit obtained from high - level context needs  to be figured out.  Acknowledgment  The research leading to the presented results has be en  undertaken within the SWARMs   European project (Smart and  Networking   Underwater   Robots   in   Cooperation   Me shes),  under Grant Agreement n. 662107 - SWARMs - ECSEL - 2014 - 1,  partially supported by the ECSEL JU and the Spanish Ministry  of Economy and Competitiveness (Ref: PCIN - 2014 - 022 - C02 -  02).  References  [1]   C. Perera, A.   Zaslavsky, P. Christen, and D. Georgakopoulos, “ Context -  Aware   Computing   for   The   Internet   of   Things:   A   Survey,”   IEEE  Commun. Surv. Tutor. , vol. 16, no. 1, pp. 414 – 454, 2014.  [2]   X. Li, J. - F. Martínez, J. Rodríguez - Molina, and N. Martínez, “A Survey  on Intermedi ation Architectures for Underwater Robotics,”   Sensors , vol.  16, no. 2, p. 190, Feb. 2016.  [3]   K. B. Laskey, “MEBN: A language for first - order Bayesian knowledge  bases,”   Artif. Intell. , vol. 172, no. 2 – 3, pp. 140 – 178, Feb. 2008.  [4]   R. N. Carvalho, S. Matsumoto, K.   B. Laskey, P. C. G. Costa, M.  Ladeira , and L. L. Santos, “Probabilistic Ontology and Knowledge  Fusion for Procurement Fraud Detection in Brazil,” in   Uncertainty  Reasoning for the Semantic Web II , vol. 7123, F. Bobillo, P. C. G.  Costa,   C.   d’Amato,   N.   Faniz zi,   K.   B.   Laskey,   K.   J.   Laskey,   T.  Lukasiewicz,   M.   Nickles,   and   M.   Pool,   Eds.   Berlin,   Heidelberg:  Springer Berlin Heidelberg, 2013, pp. 19 – 40.  [5]   R. N. Carvalho, R. Haberlin, P. C. G. Costa, K. B. Laskey, and K.  Chang,   “Modeling   a   Probabilistic   Ontology   for   M aritime   Domain  Awareness,” presented at the 2011 Proceedings of the 14th International  Conference on Information Fusion (FUSION), Chicago, USA, 2011, pp.  1 – 8.  [6]   G.   Chantas   et al. , “Multi - Entity Bayesian Networks for Knowledge -  Driven Analysis of ICH Content,”   in   Computer Vision   -   ECCV 2014  Workshops , vol. 8926, L. Agapito, M. M. Bronstein, and C. Rother, Eds.  Cham: Springer International Publishing, 2015, pp. 355 – 369.  [7]   The UnBBayes tool. Accessible online:   http://unbba yes.sourceforge.net