1 / 10

Exam Format

Exam Format. 4 questions You must do all questions There is choice within some of the questions Learning Outcomes: Explain the main approaches in classical planning as well as recent methods of planning Understand and discuss the advantages and limitations of these approaches

fallon
Download Presentation

Exam Format

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Exam Format • 4 questions • You must do all questions • There is choice within some of the questions • Learning Outcomes: • Explain the main approaches in classical planning as well as recent methods of planning • Understand and discuss the advantages and limitations of these approaches • Apply the presented planning algorithms to unseen examples

  2. Questions • Question 1 • Short essay discussing different planning technologies and their advantages and disadvantages • LO2: Comparing different planners and discussing their strengths and weaknesses • Questions 2 and 3 • Apply the planning algorithms we studied to a particular problem • LO3: apply an algorithm to an unseen example • You will have some choice about which algorithms to apply • Do read the question carefully!

  3. Questions, ctd. • Question 4 • Short answers discussing particular planning algorithms and how they work • You’ll need to know about things like the Sussman anomaly, mutexes, partial order planning, etc • Not details of the algorithms, but the ideas that are behind them • LO1: explain a planning algorithm

  4. Algorithms You Should Know • The situation calculus • Frame problem • STRIPS • Sussmann anomaly • Non-Linear planning / Plan-space planning • Graphplan • SATPlan • The Davis-Putnam procedure • How to translate a planning problem into a SAT problem • Hierarchical Planning • Situation abstraction • Operator abstraction • HTNs

  5. 2006 Exam • Describe the difference between a state-space and plan-space planner. What are their strengths and weaknesses? What kinds of problems (if any) can one solve but not the other? • State-space planners build plans by searching in the graph of the state-transition system. Plan-space planners build plans by incrementally adding actions or constraints to a partial plan • Because of this, a solution is a partial order for which all consistent total orders are executable plans • Plan-space planners separate the actions in a plan from the order they are done in, so the control strategy of the planner has less influence on the kinds of plans that can be generated • The plan space makes the reasons for plan structure explicit, so explanation is easier • The state space is typically assumed to be finite, while the plan space is not • The search space is simpler, so state-space planners are often faster, and scale better • Both problems can in theory solve the same set of problems - the differences come from the specific search algorithms of planners, not from the state-space/plan-space approach. (2)

  6. 2006 Exam • Hierarchical Task Networks (HTNs) are another planning approach that has been quite successful in some problems. Describe HTN planning, how it differs from classical planning approaches such as STRIPS, and suggest (with reasons) some domains in which it might be particularly appropriate. • In classical planning the objective is to find a set of actions from a start state to a goal state. In HTN planning it is rather to provide a way of achieving a task or set of tasks. Tasks are divided into primitive tasks that can be executed directly, and non-primitive tasks which must be decomposed. Task decomposition is performed using a library of methods, each of which is a way to turn a more abstract task into a sequence of simpler tasks. Thus HTN planning differs from classical planners in that they are hierarchical rather than working in a flat space, and in the fact that the methods encode very specific information about solving problems, so the planning problem contains far more domain knowledge. HTN planners have been successful in areas where domain experts can fairly easily encode what needs to be done in terms of ``recipes'' to follow. A good example is in computer games.

  7. 2006 Exam: Domain

  8. 2006 Exam • Above is a planning domain involving driving vehicles and travelling between locations. Show in detail how STRIPS solves this problem. Assume optimal choices. • Discuss the reasons why STRIPS is incomplete. Use examples to illustrate problems STRIPS cannot solve, and explain why not • STRIPS only searches for actions to achieve the preconditions of the last operator added to the plan. This means that interleaving actions from different parts of the plan is impossible. For example, the switching variable values problem is unsolvable by STRIPS because it will make the first assignment, commit to the plan to solve that subgoal, and then be unable to make the second assignment as the value it needs to move has already been overwritten

  9. 2006 Exam • Explain the approach used in planning as satisfiability algorithms such as SATPLAN. Contrast it with the approach in the situation calculus. • The idea in SATplan is to define the initial and goal state, as well as all the operators in propositional logic, and use efficient SAT solvers to find a model of all formulae which then corresponds to a plan. SitCalc uses first order logic rather than propositional logic, and solves the problem by deductive reasoning, rather than simply finding an assignment of values to variables that satisfies the formula • Show how the initial state and goal in the driving domain above could be represented as a satisfiability problem. How are the planning operators represented? • At(Jack, here,0)  At(Bulldozer, here,0)  Vehicle(Bulldozer,0)  Mobile(Jack, 0)  Person(Jack,0) ...  At(Bulldozer, there,1)  At(Jack, here, 1) • In rules, all variables are instantiated by all possible values up to a particular evaluation depth, so all actions can be applied at any point in the plan. Each rule is an implication (if this is true at time t, then this is true at t+1)

  10. 2006 Exam • Draw a plan graph for this problem, showing all the mutexes, and the final plan found. Note that you do not need to show all the smaller plan graphs created before the final one. • This version of the problem is unsolvable. Show the plan graph for this problem and explain how GraphPlan knows that there is no solution. • What would STRIPS do on the problem in Part (b)?

More Related