Outline for 4/23
Logistics
Course Topics by Week
Agent vs. Environment
Two Approaches to Agent Control
Planner
Input Representation
Representing Actions
How Represent Actions?
STRIPS Actions
Action Schemata
Planning as Search
Forward-Chaining World-Space Search
Backward-Chaining World-Space Search
Planning as Search 2
Plan-Space Search
Planning as Search 3
Planning Graph
Constructing the planning graph…
Mutual Exclusion
Graphplan
Searching for a Solution
Dinner Date
Are there any exclusions?
Do we have a solution?
Extend the Planning Graph
One (of 4) possibilities
Summary Planning
Outline
Immobile Robots Example 2: Cassini Saturn Mission
Solution: Part 1 Model-based Programming
Solution: Part 2 Model-based Deductive Executive
Solution: Part 3 Risc-like Best-first, Deductive Kernel
Consider a sub family of model-based optimal controllers where...
A family of increasingly powerful deductive model-based optimal controllers
Specifying a valve
Mode identification + reconfiguration
Example: Cassini propulsion system
MI/MR as combinatorial optimization
Generic LTMS interface
Using the LTMS in MI and MR
Plain Propositional Logic
Transition systems
Markov models
Concurrent transition systems
Trajectories of concurrent transition systems
Transition system models
Specifying transition systems
Specifying a valve transition system
Valve transition system (cont.)
Configuration Manager
Mode Identification
Characterizing MI
Mode Reconfiguration: Reachability in the next state
Characterizing MR
Statistically Optimal Configuration Management
MI and MR as combinatorial optimization
Limitations of Model-based Configuration Management
Limitations (cont.)
Model-based Reactive Executive
Model-based Reactive Control System < S, ???????
Why plan myopically by assuming the most likely state is correct?
How does MR search efficiently over the set of feasible target states?
Model-based Reactive Planning Relation to STRIPS planning?
Comparing MRP and STRIPS
How Burton Achieves Reactivity
Driver Valve Example
Model Compilation
Compile by Generating Prime Implicates
Simplifying to Strips
Simplifying to Strips (cont.)
Reasons Search is Needed
Exploiting Causality to Avoid Threats
Avoiding Clobbering Sibling Goals
Burton: Online Algorithm (partial)
Exploiting Safety
Avoiding Deadend (Sub) Goals
Defining Reversibility
Reversibility Labeling Algorithm
Burton: Online Algorithm
Incorporating Repair Actions
Eliminating Cost of Finding Transition Paths: Generating Concurrent Policies
Burton computes next action (step 1)
Burton computes next action (step 2)
Failure occurs during plan execution Burton computes next action - step 3
Burton computes next action (step 4)
Complexity: Constant Average Cost
Autonomous System Coding Challenge
Concurrent Transition Systems
Representation with Modal Logic
Demonstration of Model-based Autonomy Capabilities
Future: Systems that Model & Adapt
Future: Systems that Seek Information
Future: Systems that Anticipate
Email: weld@cs.washington.edu
Other information: CSE 592, Lecture 4
Download presentation source