CS541 Review Jim Blythe 2 Planning for the Grid USC INFORMATION - - PDF document

CS541 Review

Jim Blythe

2

USC INFORMATION SCIENCES INSTITUTE

Planning for the Grid

3

USC INFORMATION SCIENCES INSTITUTE

LIGO’s Pulsar Search (Laser Interferometer Gravitational-wave Observatory)

archive

I nterferom eter Long tim e fram es Store raw channels Short tim e fram es

Hz Time

Single Fram e Extract channel transpose Tim e- frequency I m age Find Candidate event DB Short Fourier Transform 3 0 m inutes Extract frequency range Construct image

4

USC INFORMATION SCIENCES INSTITUTE

Operators include data dependencies, host and resource constraints.

(operator pulsar-search (preconds ((<host> (or Condor-pool Mpi)) (<start-time> Number) (<channel> Channel) (<fcenter> Number) (<right-ascension> Number) (<sample-rate> Number) (<file> File-Handle) ;; These two are parameters for the frequency-extract. (<f0> (and Number (get-low-freq-from-center-and-band <fcenter> <fband>))) (<fN> (and Number (get-high-freq-from-center-and-band <fcenter> <fband>))) (<run-time> (and Number (estimate-pulsar-search-run-time <start-time> <end-time> <sample-rate> <f0> <fN> <host> <run-time>))) …) (and (available pulsar-search <host>) (forall ((<sub-sft-file-group> (and File-Group-Handle (gen-sub-sft-range-for-pulsar-search <f0> <fN> <start-time> <end-time> <sub-sft-file-group>)))) (and (sub-sft-group <start-time> <end-time> <channel> <instrument> <format> <f0> <fN> <sample-rate> <sub-sft-file-group>) (at <sub-sft-file-group> <host>))))) (effects () ( (add (created <file>)) (add (at <file> <host>)) (add (pulsar <start-time> <end-time> <channel> <instrument> <format> <fcenter> <fband> <fderv1> <fderv2> <fderv3> <fderv4> <fderv5> <right-ascension> <declination> <sample-rate> <file>)) ) ))

5

USC INFORMATION SCIENCES INSTITUTE

Grid Grid Grid

workflow executor (DAGman)

Execution Workflow Planning

Globus Replica Location Service Globus Monitoring and Discovery Service

Information and Models

Metadata Catalog Service Resource Models

detector

Raw data

Concrete Workflow tasks

Models and current state information Monitoring information High-level specs of desired results and intermediate data products

Dynamic information

Request Manager Current State Generator Submission and Monitoring System

AI-based Planner

6

USC INFORMATION SCIENCES INSTITUTE

Temporal logics for planning

7

USC INFORMATION SCIENCES INSTITUTE

Fahiem Bacchus

8

USC INFORMATION SCIENCES INSTITUTE

Fahiem Bacchus

9

USC INFORMATION SCIENCES INSTITUTE

Heuristic search planning

10

USC INFORMATION SCIENCES INSTITUTE

Derive cost estimate from a relaxed planning problem

Ignore the deletes on actions BUT – still NP-hard, so approximate: For individual propositions p:

d(s, p) = 0 if p is true in s = 1 + min(d(s, pre(a))) otherwise [min over actions a that add p]

11

USC INFORMATION SCIENCES INSTITUTE

HSP2 overview

Best-first search, using h+ Based on WA* - weighted A*:

f(n) = g(n) + W * h(n). If W = 1, it’s A* (with admissible h). If W > 1, it’s a little greedy – generally finds solutions faster, but not optimal.

In HSP2, W = 5

12

USC INFORMATION SCIENCES INSTITUTE

HSPr problem space

States are sets of atoms (correspond to sets of states

in original space)

initial state is the goal G Goal states are those that are true in s0 (initial state

in planning problem)

Still use h+. h+(s) = sum g(s0, p)

13

USC INFORMATION SCIENCES INSTITUTE

Mutexes in HSPr, take 2

Better definition:

A set M of pairs R = {p, q} is a mutex set if (1) R is not true in s0 (2) for every op o that adds p, either o deletes q

• r o does not add q, and for some precond r of o,

{r, q} is in M. Recursive definition allows for some interaction of the

• perators

14

USC INFORMATION SCIENCES INSTITUTE

Temporal reasoning and scheduling

15

USC INFORMATION SCIENCES INSTITUTE

Temporal planning with mutual exclusion relation

Propositions and actions are monotonically

increasing, no-goods monotonically decreasing:

16

USC INFORMATION SCIENCES INSTITUTE

ASPEN

Combine planning and scheduling steps as

alternative ‘conflict repair’ operations

Activities have start time, end time, duration Maintain ‘most-commitment’ approach – easier to

reason about temporal dependencies with full information

C.f. TLPlan

17

USC INFORMATION SCIENCES INSTITUTE

Contributors for a non-depletable resource violation

18

USC INFORMATION SCIENCES INSTITUTE

Contributors for a depletable resource violation

19

USC INFORMATION SCIENCES INSTITUTE

Learning search control knowledge and case-based planning

20

USC INFORMATION SCIENCES INSTITUTE

Using EBL to improve plan quality

Given: planning domain, evaluation function

planner’s plan, a better plan

Learn: control knowledge to produce the better plan Explanation used: explain why the alternative plan is

better

Target concept: control rules that make choices

based on the planner state and meta-state

21

USC INFORMATION SCIENCES INSTITUTE

Architecture of Quality system

22

USC INFORMATION SCIENCES INSTITUTE

Explaining better plans recursively: target concept: shared subgoal

23

USC INFORMATION SCIENCES INSTITUTE

Hamlet: blame assignment

24

USC INFORMATION SCIENCES INSTITUTE

Probabilistic planning

25

USC INFORMATION SCIENCES INSTITUTE

Sources of uncertainty

Incomplete knowledge of the world (uncertain initial

state)

Non-deterministic effects of actions Effects of external agents or state dynamics.

26

USC INFORMATION SCIENCES INSTITUTE

Dealing with uncertainty: re-planning and conditional planning

Re-planning:

Make a plan assuming nothing bad will happen Build a new plan if a problem is found (either re-plan to the goal state or try to repair the plan) In some cases, this is too late.

Deal with contingencies (plans for bad outcomes) at planning

time, before they occur.

Can’t plan for every contingency, so need to prioritize Implies sensing

Build a plan that reduces the number of contingencies requires

(conformant planning)

May not be possible

27

USC INFORMATION SCIENCES INSTITUTE

A Buridan plan based on SNLP

28

USC INFORMATION SCIENCES INSTITUTE

Computing the probability of success 2: Bayes nets

Time-stamped literal node Action outcome node

What is the worst-case time complexity

• f this

algorithm?

29

USC INFORMATION SCIENCES INSTITUTE

MAXPLAN

Inspired by SATPLAN. Compile planning problem to an instance

• f E-MAJSAT

E-MAJSAT: given a boolean formula with variables that are

either choice variables or chance variables, find an assignment to the choice variables that maximizes the probability that the formula is true.

Choice variables: we can control them

e.g. which action to use

Chance variables: we cannot control them

e.g. the weather, the outcome of each action, ..

Then use standard algorithm to compute and maximize

probability of success

30

USC INFORMATION SCIENCES INSTITUTE

Probabilistic planning: exogenous events

31

USC INFORMATION SCIENCES INSTITUTE

Representing external sources of change

Model actions that external agents can take in the same way as actions that the planner can take. (event oil-spills (probability 0.1) (preconds (and (oil-in-tanker <sea-sector>) (poor-weather <sea-sector>))) (effects (del (oil-in-tanker <sea-sector>)) (add (oil-in-sea <sea-sector>))))

32

USC INFORMATION SCIENCES INSTITUTE

Computing the probability of success using a Bayes net

33

USC INFORMATION SCIENCES INSTITUTE

Example: the weather events and the corresponding markov chain

The markov chain shows possible states independent of time. As long as transition probabilities are independent of time, the

probability of the state at some future time t can be computed in logarithmic time complexity in t.

The computation time is polynomial in the number of states in

the markov chain.

34

USC INFORMATION SCIENCES INSTITUTE

The event graph

Captures the dependencies between events needed to build

small but correct markov chains.

Any event whose literals should be included will be an ancestor

• f the events governing objective literals.

35

USC INFORMATION SCIENCES INSTITUTE

Probabilistic planning: structured policy iteration

Craig Boutilier

36

USC INFORMATION SCIENCES INSTITUTE

Structured representation

States decomposable into state variables Structured representations the norm in AI

STRIPS, Sit-Calc., Bayesian networks, etc. Describe how actions affect/depend on features Natural, concise, can be exploited computationally

Same ideas can be used for MDPs

actions, rewards, policies, value functions, etc. dynamic Bayes nets [DeanKanazawa89,BouDeaGol95] decision trees and diagrams [BouDeaGol95,Hoeyetal99]

Craig Boutilier

37

USC INFORMATION SCIENCES INSTITUTE

Action Representation – DBN/ADD

Craig Boutilier

38

USC INFORMATION SCIENCES INSTITUTE

Structured Policy and Value Function