Planning to Learn, Learning to Plan

1
Planning to Learn, Learning to Plan
2
Announcements

• Quiz
• A Review of AI Planning Techniques
• Reading for next time Cognitive model for
  planning
• Allegro for Windows
• Project deadline I

3
What is planning?

• Figuring out what to do next
• Wumpus agent already does that with
• First-order logic
• Resolution
• Shortcomings
• Default values
• Efficiency

4
Why do we need planners?
5
STRIPS Planning

• State space search
• Just like the search we saw last quarter
• Its all in the operators
• What does a STRIPS operator look like?

6
STRIPS Operators
At(here), Path(here, there)
Go(there)
At(there), ?At(here)
7
Planning Terminology I

• STRIPS ops
• Action description
• Precondition
• Effect / Postconditions / Add Delete
• Operator schemata
• When is operator o applicable in situation s?

8
Planning Terminology II

• The final frontier of planning space
• State (situation) space
• Plan space
• Plan space is populated by __________
• Operators
• Refine by eliminating plans from the set of plans
  under consideration
• Modify plans by messing with them

9
Planning in Plan Space

• NOAH planner (Sacerdoti 1975) was first
  partial-order planner
• In state space, solution is a path
• Series of operators
• In plan space, series of plan transformations
• Examples Expand detail, adding ordering
  constraints

10
Pruning the Search Space

• Cutting down the search space
• Means-end analysis
• Prioritize goals
• Identify interactions
• Parallelism
• Abstraction levels
• Different approaches
• Early NOAH, ABSTRIPS (Sacerdoti 1973)

11
A Problem in Plan Space

• The goal Getting milk, banana and a drill and
  heading home
• Actions
• Go From here to there
• Buy Weve got money
• Good things to know
• Hardware stores sell drills
• Supermarkets sell milk and bananas

12
Getting Started

• Start with an initial plan

Start
At(Home) Sells(SM, Banana)
Sells(SM, Milk) Sells(HWS, Drill)
Have(Milk) Have(Banana)
Have(Drill) At(Home)
Finish
13
Next Step
Start
At(s) Sells(s, Bananas)
At(s) Sells(s, Milk)
At(s) Sells(s, Drill)
Buy(Milk)
Buy(Bananas)
Buy(Drill)
Have(Drill) Have(Milk) At(Home) Have(Banana)
Finish
14
What have we got?

• Protection
• Need to have drill
• Buy drill achieves Have(Drill)
• If we mess with drill buying, then
• When doesnt it matter?

15
And after that...
Start
At(SM) Sells(SM, Bananas)
At(HWS) Sells(HWS, Drill)
At(SM) Sells(SM, Milk)
Buy(Milk)
Buy(Bananas)
Buy(Drill)
Have(Drill) Have(Milk) At(Home) Have(Banana)
Finish
16
Whats the problem?

• Need to
• Go from home to hardware store
• Go from home to supermarket
• Pick one and then...

17
Interactive Problems

• Big red arrows are protected links
• Protected from threats
• Change the ordering
• The problem in the abstract
• Suppose S1 achieves c for S2
• Now S3 comes along and clobbers c

18
Project topics

• Planning
• Build a planner from scratch (AB)STRIPS, NOAH
• Explore current planners
• Robotics
• Investigate reactive planning write a series of
  RAPs
• Build a robot using subsumption

19
Project Topics II

• Perception
• Explore audio perception
• Write an object recognizer (Pepsi cans in Waynes
  World)
• Machine learning
• Look at data mining (name email from newsgroup
  sigs)

20
Project Topics III

• Uncertainty
• Build a system that constructs Bayesian networks
• Look at HMMs in speech recognition
• Natural language
• Write a story generator
• Tell jokes