A Risk-Driven Approach for Goal Deliberation

1
A Risk-Driven Approach for Goal Deliberation

• Yudistira Asnar, Paolo Giorgini, Nicola Zannone

2
Outline

• Motivation
• Case Study
• BDI Architecture
• Goal-Risk Framework
• Jadex an Implementation of BDI Architecture
• Realizaation GR Framework in Jadex
• Concluding Remarks

3
Definition

• An agent is anything that can be viewed as
  perceiving its environment through sensors and
  acting upon that environment through effectors
• Rational agent should do the action to achieve
  its goals with maximized its performance measure
• evidence, past experience, or belief
• built-in knowledge
• An intelligent agent is one that is capable of
  flexible autonomous action in order to achieve
  its goals
• reactivity able to perceive their environment
  and response it
• pro-activeness able to exhibit goal-directed
  behavior by taking an initiative
• social ability capable to interact with other
  agents.

4
Motivation

• Agent has proved to be useful in critical systems
  because of
• Simple to Model Agent, in terms of its Mental
  States
• Belief-Desire-Intention (BDI) Rao 91
• Belief-Choice-Capability-Obligation Shoam 93
• Behave Autonomously
• Deliberation process
• In a safety-critical mission, an agent can
  replace a human instead
• E.g., reconnaissance, combat-mission

5
Case Study - Unmanned Aerial Vehicle

• An aircraft without pilot, that either is
  controlled remotely or flies autonomously
• Used for decoy, reconnaissance, combat, and even
  for research and civil purposes
• Attempts to use intelligent agents in managing
  UAV
• avoid the risk of human life loss
• response event autonomously w/o waiting the
  ground control commands
• take countermeasures to minimize malicious events
• A human pilot can react-learn instantaneously and
  also anticipate future events

6
Objectives

• Integrate Goal Risk analysis into an agent
  implementation platform
• an agent can reason and anticipate future events
• Requirements
• Able to model agent mental states
• E.g., BDI model
• Able to perform reasoning
• Standard Agent Platform

7
BDI Architecture

• BDI concept Bratmann 87 is well recognized in
  the philosophical and AI to model behaviors of a
  rational agent
• Adapted into a formal model such that is more
  suitable for the further implementation in terms
  of software agents Rao 91
• Belief
• information of the state of the world that is
  believed to be true
• Desire (or Goal)
• state of the world that an agent is trying to
  achieve
• Intention
• means to achieve the agent's desires (i.e.,
  represented as plans)

8
BDI Architecture (2)

• Deliberation process is realized by a Meta-Level
  reasoning
• It is used to choose the most proper means to
  achieve a goal
• Belief-based
• Predefined-rules

Rao 95
9
Tropos Goal Risk Framework

• Three layers of analysis
• Goal layer
• Strategic interests of an agent
• Goals and Plans
• Event layer
• Uncertain circumstance that effects goal layer
• Events
• Treatment layer
• To treat the effects of event layer
• Treatments
• Relations
• AND/OR Decomposition
• Contribution
• Means-end

10
GR Reasoning

• Define alternative solutions
• E.g., A set of leaf goals
• Define plans to achieve each alternative
• E.g., a set of plans to achieve leaf goals
• Assess the risk (also cost) of each alternative
• If the risk (also)is unacceptable
• Introduce treatments as part of alternative
  solution
• Back to (3)
• otherwise
• Add alternative (treatment) as a solution

11
Jadex

• An agent middleware platform and a reasoning
  engine
• Agent Middleware Platform It is used JADE or
  DIET as host platform
• common services
• message passing
• agent persistency
• agent management
• standards e. g. FIPA conform
• often assume simple task-based agent models
• Reasoning Engine
• focus rational agents
• agent architectures
• programming languages
• based on theories from other disciplines
• Philosophy (BDI)
• Biology (subsumption)

Taking from Jan Sudeikiats Presentation
12
Jadex vs Tropos

• Both of them have
• Belief
• --------
• Means-End relation
• Goal
• Plan

BUT .
13
Tropos vs Jadex

• Tropos
• Goal Type
• Achieve
• Avoid
• Maintain
• Plan
• achieve a goal
• Means-End
• Define means to achieve leaf goals

• Jadex
• Goal Type
• Achieve
• Maintain
• Perform
• Query
• Plan
• achieve a goal
• trigger subgoals
• Means-End
• Define means to handle all goals

14
Jadex Agent Architecture
Different from the one in GR-Model
Meta-Level Reasoning
Pokhar 05
15
Goal Cycle in Jadex
Pokhar 05
16
Jadex vs Tropos

• Differences
• AND-Decomposition
• Tropos ? Knowledge Level
• An UAV agent knows that to investigate enemy
  area, it needs to 1) define the flight airways,
  2) fly to the target, 3) Identify the target,
• Jadex ? Plan Level
• To investigate enemy area-G1, an UAV has a plan
  that triggers subgoals which are 1) define the
  flight airways-G4, 2) fly to the target-G5, 3)
  Identify the target-G6,
• OR-Decomposition
• Tropos ? Knowledge level
• To take a picture-G7, an UAV can apply 1)
  take-store schema-G8 or 2) take-transmit
  schema-G9
• not supported in Jadex
• Contribution
• not supported in Jadex

17
GR Framework Realization

• Adapt the differences
• How to mimic decomposition relations
• How to represent contribution relations such that
  it is considered during goal reasoning
• Deliberation Process
• If there are several alternatives, an agent
  should choose among alternatives deliberately
  considering the risks
• Reasoning (i.e., risk reasoning)
• Adapt the GR reasoning in the deliberation
  process

18
Realization Rules Thumb

• AND-Decomposition
• Define a plan that triggers the subgoals when the
  top goal is activated by an agent

19
Realization Rules Thumb

• OR-Decomposition
• Define plans that each plan will trigger a
  subgoal
• Reasoning chooses only one plan among the
  applicable ones

20
Realization Rules Thumb

• Contribution Relations from the Treatment Layer
• A goal/plan can be achieved/executed with or
  without the treatment
• Depends on the level of risk
• Treated as OR-Decomposition

21
Belief

• Knowledge of the agent
• Threshold of Risk (also cost)
• The Risk Level (also cost) of the chosen
  alternative
• GR Model
• All Layers
• Values of each constructs
• Contribution relations are represented in GR
  Model (i.e., belief) for reasoning purpose

22
Goal

• As default all goals in GR model are achieve goal
• An additional goal
• Maintain risk level
• The goal that has several applicable plans
• e.g., or-decomposition, contribution relations
  from the treatment layer
• Meta level reasoning (e.g., choose_planG07)

23
Plan

• Define the plan that will be executed if the
  beliefs change (e.g., values of events, plans,
  treatment)
• Define the plan to handle each goal
• Define the plan to handle meta-level reasoning of

24
GR Model in Jadex
25
Note

• GR reasoning
• Done at the beginning when an agent desires to
  achieve a particular goal
• Define the overall plans that must be executed
  considering the risks (also costs)
• Deliberation process in Jadex
• Executed each time there is an alternative
• The overall plans is not predefined

26
Concluding Remarks

• The Tropos Goal-Risk framework, which consists of
  a modeling framework and reasoning mechanisms
  implemented in Agent Platform, namely Jadex.
• The limitations in Jadex does not limit the
  expressiveness of the intended goal deliberation
  process