Quiz

1
Quiz

• The desJardins, Durfee, Ortiz and Wolverton paper
  organizes distributed planning research into two
  basic categories. Summarize the difference
  between these categories.
• The desJardins and Wolverton paper is based
  around a technical definition of the term
  irrelevance. Why do they care about
  irrelevance?

2
Issues

• How distributed?
• How much information shared
• Who benefits? The collective? The individual?
• Whos in charge?
• How do agents communicate, coordinate, negotiate?

3
Issues

• How distributive
• One planner, multiple bodies?
• Multiple mars rover problem 3 dumb rovers with
  planner at base station
• One goal, multiple planners
• Military operations (in theory) Attack beach
  naval planner, marine planner
• Multiple (maybe overlapping, maybe conflicting)
  goals
• Military operations (in practice)
• What is shared?
• One plan known by everyone
• Partially shared plan (marine/navy example)
• Completely hidden plan (adversarial planning,
  chess)

4
Issues

• Who benefits?
• Global vs. local utility function?
• Whos in charge?
• One person vs. hierarchies vs. nobody
• Centralized task/goal allocator, plan merger
• Marketplace agents that bid for tasks
• How do agents communicate/negotiate?
• Rigid protocols
• Turn taking,
• Chaos

5
DSIPE

• HTN planner based on SIPE
• More complex than UMCP
• Partial order
• Conditions on decomposition rules (filter vs.
  pre)
• Phantomization
• Task vs goal muddle
• Characteristics
• Multiple planers, partially shared goal
• Global utility function (top-level-goal
  achievement)
• Centralized task/goal allocator, plan merger
• Partially shared knowledge
• Structured protocols

6
DSIPE
A
C
B
B
C
E
D
G
F
F
D
7
What to communicate

• Cooperative agents, could use global plan but
  dont for efficiency reasons (simplify local
  plans)
• But still want to avoid bad conflicts
• Dont communicate irrelevant stuff
• So what is relevant? Use heuristics

8
Multi-agent Example

• Simple logistics domain
• Two agents, one resource (a car), differing goals
• Drive(?person, ?start, ?end)
• Pre inside(?person, car), at(car, ?start)
• Eff at(car, ?end), -at(car, ?start)
• Enter(?person, ?loc)
• Pre at(?person, ?loc), at(car, ?loc)
• Eff inside(?person, car), -at(?person, ?loc)
• Exit(?person, ?loc)
• Pre inside(?person, car), at(car,?loc)
• Eff at(?person,?loc), -inside(?person,car)
• Surf(?person)
• Pre at(?person, beach)
• Eff catch-some-waves(?person)
• Buy-lotto-ticket(?person)
• Pre at(?person, quicky-mart)
• Eff have-money(?person)

9
Multi-agent Example

• Simple logistics domain
• Two agents, one resource (a car), differing goals
• Goal have-money(you)
• Drive(?person, ?start, ?end)
• Pre inside(?person, car), at(car, ?start)
• Eff at(car, ?end), -at(car, ?start)
• Enter(?person, ?loc)
• Pre at(?person, ?loc), at(car, ?loc)
• Eff inside(?person, car), -at(?person, ?loc)
• Exit(?person, ?loc)
• Pre inside(?person, car), at(car,?loc)
• Eff at(?person,?loc), -inside(?person,car)
• Buy-lotto-ticket(?person)
• Pre at(?person, quicky-mart)
• Eff have-money(?person)

10
Multi-agent Example

• Simple logistics domain
• Two agents, one resource (a car), differing goals
• Goal catch-some-waves(?person)
• Drive(?person, ?start, ?end)
• Pre inside(?person, car), at(car, ?start)
• Eff at(car, ?end), -at(car, ?start)
• Enter(?person, ?loc)
• Pre at(?person, ?loc), at(car, ?loc)
• Eff inside(?person, car), -at(?person, ?loc)
• Exit(?person, ?loc)
• Pre inside(?person, car), at(car,?loc)
• Eff at(?person,?loc), -inside(?person,car)
• Surf(?person)
• Pre at(?person, beach)
• Eff catch-some-waves(?person)