Title: Diagnosing a team of agents: Scaling up

1
Title Diagnosing a team of agents Scaling up

• Written by Meir Kalech and Gal A. Kaminka
• Presented by Reymes Madrazo-Rivera

2
Goal

• To explain a way of dealing with disagreements
  when they occurs
• in a team environment. The authors came up with
  three different
• ways of detecting and diagnosing disagreements as
  soon as they
• happen in team of behavior-based agents. They
  were especially
• focused in reducing the communications and
  computational
• requirements of the diagnosis for teams composed
  of
• behavior-based agents.

3
Structure

• Some useful concepts
• Introduction to the topic
• Previous work in this area
• New methods for diagnosing disagreements
• Experiments and conclusions

4
Useful concepts

• Beliefs They provide a set of data describing
  the state of the environment at a given moment.
• Behavior-based agents System which are
  decomposed into different processes or behaviors,
  where each process is responsible of interacting
  with a given characteristic of the environment
  1.
• Behavior It is a mechanism of controlling the
  actions executed by the agent once certain
  preconditions, represented by predicates, are
  fulfilled. That is, given certain precondition
  that behavior can be executed. When analyzing
  behavior-based agents, for each agent is defined
  a hierarchy of behaviors, which are arranged from
  a general one to specific behaviors

5
Useful concepts (cont.)

• Behavior path Specific sequence of behaviors
  followed given that for each one of them their
  preconditions were fulfilled. For example, for
  Fig.1 could be defined paths (A B E), A C F) and
  (A D E).
• Team behavior Specific behavior path that should
  be executed at the same time for all members of
  team. For example, in Robocup could be defined
  two team behaviors, Attack and Defend, so that,
  given an initial precondition (s), all members of
  the team know that they have to follow certain
  behaviors to score a goal.

6
Introduction

• When a team is playing, it should be an agreement
  between all the teammates
• about what strategy in general to follow. That
  strategy is given by the team
• behavior. Nevertheless it could be disagreements
  between members of a team
• what is modeled by the fact that they execute
  different team behaviors. For
• example, members of a team in Robocop are in
  Attack mode, but some of them
• executed the behavior path corresponding to
  Defend mode for some reason.
• The process followed to identify disagreements
  consists of comparing the agents
• team behaviors and later identifying which
  beliefs led to that team behavior. This is
• performed by a dedicated agent, who perform a
  diagnosis task of identifying the
• team members beliefs and selecting the
  conflicting ones that led to the
• disagreement.

7
Previous work for diagnosing disagreements

• Two methods were used before this article
• Reporting Each teammate communicate its belief
  to the diagnosing agent, and the last one
  compares each other to find contradictions. It
  requires polynomial times in the number of
  agents.
• Querying It is longer process and it is the one
  that served as the base for the three new methods
  presented in this article. It has three steps
• First, all those behaviors associated to the
  computed actions (linear complexity in the number
  of behaviors) are searched by creating later a
  set of behavior-path hypotheses that contains
  behaviors associated to the observed actions.
• For each one of those behavior-path hypotheses,
  the diagnosing agent creates a set of belief
  hypotheses that are related to that specific
  behavior path hypothesis. Belief hypotheses are
  all the possible values to evaluated
  preconditions. For example, for a precondition of
  (p ? q), the possible hypotheses are (p ? q), (p
  ? ?q) and (?p ? q). (exponential complexity in
  the number of beliefs)
• With those beliefs hypotheses, the diagnosing
  agent can start querying the observed agent to
  determine if it is traversed the same behavior
  path that matches with the computed actions.
• At the end the observer agent has the beliefs for
  each observed agent and compares them (polynomial
  complexity).

8
New methods for diagnosing disagreements

• Three methods were proposed in this article. Each
  one to reduce the complexities
• mentioned before
• Behavior querying This one avoids the long
  process of deducing all possible path behaviors
  by using communication. That is, by querying the
  analyzed agent about the path that it followed.
  In that way are avoided the rest of possible
  behavior path that were not used. So now, what it
  was a process linear complexity process in the
  number of behaviors becomes in a constant
  complexity problem O(1). In fact, also will be
  reduced the number of belief hypotheses to deal
  with, because it will be carried out just for one
  behavior path instead of a group of them.

9
New methods for diagnosing disagreements (cont.)

1. Shared belief This one looks for reducing the
   exponential growth in the number of beliefs that
   is associated with a given behavior path. The
   idea now is to infer the prepositions associated
   to a belief, without considering their values.
   So, once the diagnosing agent has the
   prepositions associated to each agents
   preconditions, it compares those prepositions for
   each pair of analyzed agent to find share
   prepositions, which actually could be the source
   of contradictions. For example, consider that
   agents A and B consider prepositions (p,q) and
   (p,r) respectively. So, to determine whether A
   and B disagree, it will be enough to querying A
   and B for both values of the share preposition p
   (p and ?p). It could be that A believes p and B
   believes ?p, and that is the source of
   disagreement that caused that each one execute a
   different behavior path (Those values could be
   sensed differently because agents were in
   different physical locations). So far
   communications between agents keep increasing
   with the number of agents because all pair of
   agents are compared. But, the complexity reduce
   considerably (to a linear process in the number
   of beliefs) since inferring the exponential
   number of believe hypotheses is not necessary
   anymore.

10
New methods for diagnosing disagreements (cont.)

1. Grouping The last process carry out by the
   diagnosing agent is the comparison between
   teammates beliefs. This is a polynomial process
   in the number of agents to compare and in the
   number of beliefs involved. This last method
   creates groups of agents composed of those with
   similar role into the team and the same behavior
   path. As a consequence, the diagnosing agent will
   select a single agent from each group to compare
   their beliefs, instead of comparing the belief of
   each observed agent. This method first uses
   behavior querying to get each agent behavior
   path. Agents groups are then built by considering
   that previous information and each observed
   agents role. Then the diagnosing agent select
   one observed agent per group and carry out for
   them a share believe process to finally determine
   possible sources of disagreements between them.

11
Experiment design and valuation

• Those methods were evaluated on a multi-agent
  system (ModSAF). This
• application models teams of helicopter pilots.
  For that experiment were considered
• two roles for pilots and four behavior paths. The
  number of agents was increased
• from 6 to 150 with an step of eight.
• Four models were considered for the study in
  addition to the initial methods of
• Reporting and Querying
• Behavior The diagnosing agent just uses behavior
  querying (so, it works with the proper behavior
  path). So, it completes the process by generating
  belief hypotheses and querying observed agents
  about those belief hypotheses.
• Belief The share belief method is used to
  generate belief hypotheses (so, it works with
  values of preconditions prepositions shared by
  agents).
• BehaviorBelief Its a combination of the two
  first. After determining the behavior path of a
  given agent by using behavior querying, then
  possible disagreement is identified by using
  share belief.
• Grouping It is also a combination of the two
  first method, but this time groups of agents
  according to their role and behavior paths are
  created, and after that the share belief is
  applied to diagnose disagreements, but the
  comparison will be carried out just between the
  representative agents of each group.

12
Experiment design and valuation (cont)

• The experiment consisted of plotting each model
  in two different graphs. One of
• (Number of Beliefs Messages Used vs Number of
  Agents) and the other one of
• (Runtime in Milliseconds vs Number of Agents).
• The conclusion for that test was that the runtime
  grows polynomially in the number
• of agents for all the models due to the number of
  comparisons, except for the
• Grouping one, where the complexity is reduced to
  linear growth because we are
• dealing with a fixed number of comparisons. That
  reduced number of comparisons
• also leads to a reduction in the number of
  messages. The two other new method
• by themselves (behavior querying and share
  belief) do not seem to have much
• contribution in the reduction of runtime or the
  number of messages. That happened
• because in that experiment the authors were
  focused in checking the influence of
• the number of agents and that is why the number
  of beliefs were small as well as
• the number of behaviors.

13
Conclusions

• It was presented a method that reduces the
  communication and
• run-time when diagnosing disagreements between
  teammates
• when a great deal of agents is involved. Firstly
  was reduced the
• number of communications initially by querying
  about the behavior
• path followed. Later on the communication was
  much more
• reduced by creating groups of agents according to
  their role and
• their computed behavior path. Consequently, the
  diagnosis
• process continues by considering a representative
  agent of each
• group.

14
THANKS