Article Directory
This paper aims made it clear that:
1) planning and problem definition (PDDL) is a search problem
2) forward search, backward search, heuristic search and
3) to obtain heuristic rules and extracts from plans
Planning Problem Definition (PDDL) is a search problem
Planning Domain Definition Language (referred PDDL, planning area defined language)
Planning : design an action plan to achieve the goal
ch03 Agent and solving logic problems Agent ch07 based search is a plan Agent.
With PDDL able to describe planning issues ch03 and ch07 impossible to describe.
Therefore, the researchers plan to select the elements of (factored representation) - representation of a representation of the state of the world with a set of variables. It can be used PDDL language.
The planning problem is defined as a search problem (elements of what?)
- Status : Flow (FLUENT) (synonym state variable) conjunctions , these streams are primitives (no variables), no function atoms.
For example, the state may represent bad Poor∧Unknown the Agent;
state in a parcel delivery problems may At (Truck 1, Melbourne) ∧At (Truck 2, Sydney)
in a state, the flow is not allowed: : At (x, y) (with a variable), ¬Poor (), At ( Father (Fred), Sydney) ( using the notation function) - Operation mode : the operation mode can be used to describe the operation, the operation mode is implicitly described ACTIONS (s) and RESULT (s, a).
For example helicopter from a location to another mode of action:
Action (Fly (the p-, from, to),
PRECOND: At (the p-, from) ∧Plane§ ∧Airport (from)
∧Airport (to)
the EFFECT: ¬At (p, from) ∧At (p , to))
this mode has the action name, all variables used in list mode, the premise (precondition), and the effect (effect).
PDDL according to the results of what has changed to describe an action; not to mention things remain the same. Operation and effect are the premise text (sentence or a negative atom statement atoms) conjunctive . Premise defines the state of the operation can be performed, the effect of the definition of the results of the implementation of this action.
Obtained for all primitives operated by substituting the value of variables:
the Action (Fly (P. 1, SFO, JFK),
PRECOND: At (P. 1, SFO) ∧Plane (P. 1) ∧Airport (SFO)
∧Airport (JFK)
EFFECT: ¬At (P 1, SFO ) ∧At (P 1, JFK))
ACTIONS (s): s an action can be performed in a state, if s implies a premise. Implication also be used to express a set of semantics: s╞q text if and only if q and n are in s, and q is the negative character is not in s. In the form of symbols, it can be said a∈ACTIONS (s)) ⇔ s╞ PRECOND (a)
if the state s meet premise, we call in a state s action is applicable (applicable).
RESULT (s, a): In the definition of the result state s perform action a state s', which is represented by a set of flow (fluent), the set flow begins s, remove the stream operation effects a negative character appears ( delete the list or delete list called DEL (a)), and appears to increase the flow in the operation of the positive effects of text (referred to increase or add list listing the ADD (a)):
S '= the RESULT (S, a) = (s - DEL (a) ) ∪ADD (a)
The initial state is conjunctive primitive atoms. (For all states, the use of a closed world assumption, which means that any atom not mentioned are false.)
Target as the antecedent: Text (positive or negative character text) that may contain variables conjunction, as At (p, SFO) ∧Plane§. Quantified variables exist
when we can find a sequence of actions that s the end of the implication of the target state, the problem has been resolved. For example, the state of Plane (Plane 1) ∧At (Plane 1, SFO) implies the target At (p, SFO) ∧Plane§.
Classic planning complexity
This section considers the complexity of the planning theory, and distinguish between the two decision-making problems.
PlanSAT: asked whether a plan to solve the problem of the existence of a plan
bounded (bounded) PlanSAT: ask whether there is a plan to find the optimal solution of a length less than k.
Delimited (bounded) PlanSAT NP- is complete, the PlanSAT belongs P . In other words, the best plan is usually difficult, but sometimes sub-optimal planning is easy.
State-space search algorithm planning
The former state-space search
Description planning defines a search problem, heuristic search algorithm (Chapter 3) or a local search algorithm (Chapter 4) programming problems can be solved.
Forward Start searching state space from the initial state to find a target.
s' = RESULT (s, a ) = (s -DEL (a)) ∪ADD (a)
Forward search : from the initial state, the use of the action, the search target state forward
backward search : From the state of the target set, use the reverse action, search backwards initial state.
1) before the search to easily explore unrelated action
2) planning problems often have large state space
for forward search, it is clear that there is no accurate heuristics, even a relatively small problem instance was hopeless. But many real-world applications can be automatically exported or planning a very strong domain-independent heuristic knowledge, which makes pre-feasibility to Soso.
After searching the relevant state
From the start target application actions back, until it finds a sequence of steps to achieve the initial state. At each step of considering a set of related state, rather than a single state.
Start from the target, a set of states for the text described in the conjunctive.
Fortunately, the design PDDL represented backward movement makes it easy.
The last question is: decide what action is to retreat candidate action, we have chosen to apply the previous action - in the planning of those actions might be the next step in the backward direction, we need the relevant action - leading to the current planning target states can be used as the last step of those actions.
An action to be associated with a destination, it must contribute significantly to the goal.
Although the former is lower than the branching factor to the search, however, after making it to the fact that the use of state set instead of a single state is more difficult to come up with a good heuristic knowledge, this is the main reason for the current mainstream of the forward bias.
Heuristic planning
Without a good heuristic function, both backwards and forwards is not efficient.
Admissible heuristic. Relaxation can be derived by solving problems more easily.
The search problem like into a view state for the node, the edge of the operation. The problem is that looking for a path connecting the initial state to the target state.
There are two methods to relax this problem:
1) adding more sides, such that the path is more likely to be found
2) a combination of the plurality of nodes together, the abstract state space form state having less
ignored premise heuristic
Ignore List Delete heuristic
key idea is the definition of heuristic decomposition (decomposition)
Information sharing
Experiment download:
https://github.com/yyl424525/AI_Homework
Artificial Intelligence - A Modern Approach Chinese third edition pdf, courseware, operations and solutions, after-school exercise answers, test code and reports, over the years Kaobo title Download: HTTPS : //download.csdn.net/download/yyl424525/11310392