Introduction to CSE 591: Autonomous agents - theory and practice.

1
Introduction to CSE 591 Autonomous agents -
theory and practice.

• Chitta Baral
• Professor
• Department of Computer Sc. Engg.
• Arizona State University

2
About the course

• It is an advanced graduate level course.
• Prerequisites
• None formally, but be concerned (and discuss with
  me)
• If not taken an AI class before.
• If afraid of writing proofs.
• If you dont know what NP-completeness is.
• The course will be rigorous.

3
Grading

• Scale A, A, B, B, C, C, F.
• Grade weights
• Home work small assignments 20-30
• 2 Exams 50
• Paper 20-30
• One exam in mid-October and another on the last
  day of class (no exams on the day of finals)

4
A formal approach to design and analysis of
autonomous agents

• Lots of hype about agents.
• Used in different contexts, no uniform framework.
  
• Multiple answers to the question What is an
  agent?
• No systematic approach to design or analyze an
  agent exists.

5
Central thesis

• Give a formal approach to design and analysis of
  autonomous agents.
• Similar to the design of relational databases.
• Use this approach in different applications.
• For example network agents, embedded agents,
  diagnostic agents, remote (space) agents,
  data-consistency agents (active databases),
  viewing living cells as agents.

6
What are autonomous agents?

• Agents
• One that acts or has the power or authority to
  act.
• One empowered to act for or represent another.
• A means by which something is done or caused.
• A force or substance that causes a change.

7
What are autonomous agents? Cont.

• Autonomous
• Not controlled by others or by outside forces
• Independent in mind or judgment or government
  self directed.
• Self governing

8
Our view

• An agent is an entity (software module, piece of
  embedded hardware) that acts and causes change.
• Autonomy
• Not micro-managed
• Can take high-level directives and goals and can
  figure out what actions it needs to take and
  execute those actions.

9
What ability an autonomous agent must possess?

• Must understand high level directives
• Must have knowledge about its abilities
• What actions it can perform
• How these actions affect the world
• Must be able to reason and chalk out its plan of
  action (and revise them if things change)

10
Designing agents the steps

• Representing actions and environment.
• Objects in the environment.
• Relationship between the objects.
• Actions that the agent can perform.
• Actions that are beyond the control of the agent.
  (exogenous actions)
• How does the actions affect the environment.
• Main issue -- an appropriate language (syntax
  and semantics) for representing the above.

11
Designing agents (cont.)

• What is the goal of the agent?
• Achievement goals.
• Maintenance goals.
• In general, to make the trajectories taken by the
  agent satisfy some temporal (LTL, CTL) formula.
• Research issues Are existing temporal languages
  enough? Do we need new constructs?
• Knowledge temporal goals!
• Issues Complexity of planning with such goals.

12
Designing agents (cont)

• Agent architecture and execution languages
• Pure reactive agent a collection of rules of
  the form if observed x and context is z then
  do action y.
• The agent continuously senses the environment
  and executes the rule whose condition matches the
  sensor values.
• Research issue
• Verifying that a collection of such
  rules when executed will satisfy the goal of
  the agent.
• Automatic generation of such rules.

13
Designing agents (cont)

• Agent architecture and execution languages (cont)
• Pure deliberative agent Follows a cycle of
• Sense
• Assimilate observation (reasoning is needed here)
• Replan if necessary (reasoning and planning are
  needed here).
• Replanning becomes necessary if environmental
  actions make changes that invalidates the
  original plan.
• Execute part of the plan.
• (assimilation and planning are both hard
  NP-complete even in simple cases.)

14
Designing agents (cont)

• Agent architecture and execution languages
• Hybrid agents
• Reactive for some (important/frequent/critical)
  states and deliberative for others.
• The states for which reactions are stored are
  updated similar to the cache/main-memory
  hierarchy.
• Various levels of reactivity simplest ones spell
  out the action next use a fast (logic) program
  to decide on the next action use domain
  knowledge to do faster planning so as to come up
  with the action to be executed.

15
Designing agents -- summary

1. Actions, environment, and effect of actions.
2. Goal or directives.
3. Execution or architecture.
4. Design and analysis tasks
5. Write 1 and 2 and then,
6. Given 1 and 2 come up with 3 or
7. Given 1, 2, and 3, verify 3 is correct with
   respect to 2 or
8. Given 1, 2, and 3, and some change in 2, update 3
   to satisfy the new directives.

16
A data consistency maintenance agent. (An active
database)

• Actions and their effect on the environment
• Fluents (technical term for objects in the
  world)
• purchase(purchaseid, client, amount).
• payment(paymentid, client, amount).
• account(client, credit, status).
• Environmental actions
• insertion of tuples to the purchase and
• payment relations for existing
  clients.
• Agent's actions Any other insert, delete
  and update.

17
Active database agent (cont.)

• Actions and their effects on the environment.
  (cont.)
• insert(f) causes f
• executable insert(f) if f
• delete(f) causes f
• executable delete(f) if f
• update(f,f') causes f', f
• executable update(f,f') if f

18
Active Dtabase agent (cont.)

• Goal Make sure the stable state of the database
  satisfies the following.
• For each client c which appears in a tuple a in
  the relation account if a.credit lt 3K then
  a.status bad, and if a.credit gt 3K then
  a.status good.
• For each client c which appears in a tuple a in
  the relation account a.credit is 5K minus the
  sum of all the purchase amounts for c plus the
  sum of all the payment amounts for c.

19
Active Database agent (cont.)

• Reactive execution program (triggers)
• Trigger 1 When a tuple p is added to the
  purchase relation, then update the tuple a in
  the relation account such that p.client
  a.client so that the updated a.credit has the
  value obtained by subtracting p.amount from the
  old a.credit.
• Trigger 2 When a tuple a in the relation account
  is updated such that a.credit is less than 3K
  then update a such that a.status has the value
  bad''.
• Trigger 3 When a tuple p' is added to the
  payment relation, then update the tuple a in the
  relation account such that p'.client a.client
  so that the updated a.credit has the value
  obtained by adding p'.amount to the old a.credit.

20
Active Database agent (cont.)

• Reactive execution program (cont.)
• Trigger 4 When a tuple a in the relation account
  is updated such that a.credit is more than or
  equal to 3K then update a such that a.status has
  the value good''.
• Correctness Suppose the database is in a stable
  state (I.e., it satisfies the goal conditions)
  and an allowable exogenous action occurs, then if
  the triggers are executed they will take the
  database to a stable state.

21
Paper ideas

• New language constructs for expressing the
  ability of agents and their impact on the
  environment
• New language constructs to express goals
• Agent architecture components
• Designing and implementing an agent in a
  particular domain