Download Symposium_poster - Satyendra Singh Chouhan

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Artificial intelligence wikipedia , lookup

Transcript
Multi-Agent Planning with Joint Actions
Satyendra Singh Chouhan, Rajdeep Niyogi
Department of Computer Science & Engineering, IIT Roorkee, INDIA
[email protected], [email protected]
MAP with Joint Actions
Introduction
• Multi agent planning(MAP): “Planning by and for multiple agents”
• MAP Applications: Logistics, Supply Chain Management, Search and
Exploration, Robotics, Video games, Healthcare etc.
Joint actions in multi-agent planning raise two main issues
• Specification (input PDDL)
• Planning (Algorithm to handle joint actions scenario)
Existing Approaches to handle joint actions at specification level.
Input (MA-PDDL)
For agent α1 to αn
Output:
Sequence of actions (set)
MA-plan: (A1; A2; ….;AK)
Ai: {act(α1), act(α2) … act(αn) }
at time i
MAP system
Fig 1. Coordination between agents
Fig2. Cooperation between agents
:action push
:parameter(?a1-agent ?b –box ?l- location)
:precondition (and(at ?a1 ?l) (at ?b ?l))
:concurrent (and (push ?a2 ?b ?l)(not(= ?a1
?a2))
:effect(not (at ?b ?l))
:joint action push
:parameter(?a1-agent ?a2-agent ?b–box
?l-location)
:precondition (and(at ?a1 ?l)(at ?a2 ?l)(at
?b ?l) (not(=?a1 ?a2)))
:effect(not(at ?b ?l))
(Boutilier, C et al. 2001) [1]
(Brafman, R. et al. 2014)[2]
• The given specifications will only work when we have complete knowledge
of the domain. Most importantly, number of agents require to perform a
joint action( #2 for above push action specification).
• Above specifications will only be useful, if the total number of agents
require to perform a joint action is at most two. If there are objects that
require more than two agents then the specification will be complex.
Motivation
• MAP is a broad area of research in Artificial intelligence. Therefore, depending
on the type of problem different map approaches are used.
• Joint actions come naturally in MAP.
• There are different ways to view joint action. Consider following example to
understand it.
Example 1: Narrow Doorway Domain
A
B
• No. of agents: 2 (A, B)
• Action set {go(G), wait(W)}
• Possible joint actions would be (G, G), (G, W), (W, G) and (W, W).
However, a successful plan would be obtained when one agent waits and allows
the other agent to pass through the gate.
Example 2: Logistics Domain
Airplane1
L2
L3
Dest
City2
In this domain, at any point in time when a truck is being loaded at one location,
another truck may be unloaded at another location. An instance of joint action
would be the set of two concurrent actions represented as <load(Truck1, city1),
unload(Truck2, city2)>
Example 3: Box-pushing Domain
In box-pushing domain (Figure 2), two or more agents may need to push a heavy
box simultaneously i.e. agents have to perform the same action at the same time.
Here, an instance of a joint action would be represented as <push(Agent1, box1),
push(Agent2, box1)>
• In addition, A joint action can be a sequence of two or more actions such that a
desire effect would only be obtained if theses actions are performed in a
particular sequence.
• Existing multi-agent planners can solve MAP problems similar to example 1 and
2. However, no multi-agent exist that can solve problems similar to example 3.
www.PosterPresentations.com
MAPR
MA-FD
FMAP
√
√
√
×
√
√
×
×
×
• In [3], single agent planning approach is used to handle Map problems with
concurrent actions. In this each agent’s action is associated with subset of
objects (including agent count). Then problem is transformed into single
agent planning problem. We have performed some experiments using
above approach with some assumptions using Blackbox planner.
Sr No.
No. of
Agents
1
2
3
4
5
2
2
2
3
3
Problem
Size
(s, m, l)
(4,2,0)
(5,3,0)
(6,2,0)
(6,2,1)
(8,4,2)
Box Pushing
Domain
Solved?
Yes
Yes
Yes
Yes
No
Extended-Block
World Domain
Solved?
Yes
Yes
No
No
No
Truck2
City1
RESEARCH POSTER PRESENTATION DESIGN © 2012
Domains
Loosely-coupled planning
domains
Tightly-coupled planning
domains
Tightly-coupled planning
domains with joint actions
Current research objectives
Truck1
Source
Existing Multi agent Planning Systems
• Current state of the art multi-agent planners does not supports joint
actions.
• Designing a Distributed Multi-agent system that can solve planning
problems with joint actions.
• Analysis of complex planning domains that involve cooperation among
agents.
Future research interest
• Identifying class of problems for which proposed planning system is
complete.
• Scope of learning in Multi-agent planning.
References
1. Boutilier, C. and Brafman, R. I . 2001. Planning with concurrent interacting
actions. Journal of Artificial Intelligence Research (JAIR), pages 105–136.
2. Brafman, R. I and Zoran U. 2014. Distributed Heuristic Forward Search for
Multi-agent system, In Proceedings of 2nd ICAPS DMAP workshop, pages 1-7.
3. Crosby, M., Jonsson, A., and Rovatsos, M. 2014. A Single-Agent Approach to
Multi-agent Planning. In ECAI 2014, page 237