Competition of Distributed and Multiagent Planners (CoDMAP)

Welcome to the CoDMAP pre-registration poll. The CoDMAP competition is meant to be a preliminary version of possible future IPC Multiagent planning track. To start with something, we focus on the following target requirements:

i) deterministic model (STRIPS-compatible)
ii) domain-independent
iii) cooperative agents (common public and/or non-interfering private goals; common metrics)
iv) offline (as in IPC Deterministic Track: input → planning→ output (plan)→ check)

We know the requirements may be limiting, please feel free to state your preferred requirements in the Additional comments section.

Please, fill in the poll also if you have (currently) no planner but you are interested in the topic. If your planner has multiple configurations, please select all applicable answers and add a note to the Additional comments section. If you have multiple planners, please submit one form per planner.

It is ok if you fill in just some of the answers.

Thank you for your participation.

Contact person: michal.stolba@agents.fel.cvut.cz

Target requirements

If the planner does not sarisfy the target requirements, please specify what are the differences (e.g., “my multiagent planner is an online Dec-POMDP planner”)

(a) My planner (moreorless) satisfy the target requirements
Other:

This is a required question

Multiagent planning model

Does your planner support planning of joint-actions?

(similar as proposed in [2])

Yes
No

This is a required question

The structure of agents in my planner is:

(a) the agents are “on the same level” controlled by the planning process (described in form of algorithms and communication protocols)
(b) there is one “leader” doing most of the work supported by “participant” agents (the leader can be voted for and can change during planning)
(c) the agents form a deeper tree-like hierarchy
Other:

This is a required question

Privacy model

Pick a level of privacy best representing your planner :

(the private/public terms follow the MA-STRIPS [1] definition)

(a) no privacy - agents are allowed to communicate any information (e.g., if it helps planning efficiency)
(b) obfuscation - agents are allowed to communicate any information, but private information has to be obfuscated s.t. it is not readable by other agents (e.g., action and fact labels are replaced by hash codes)
(c) aggregation - agents are allowed to communicate any information, but private information can be communicated only in an aggregated form (e.g., a sum of costs of private actions of a private plan, merged private actions into a macro action, etc.)
(d) privacy - agents are allowed to communicate only public information
(e) total privacy - agents are allowed to communicate only public information in a minimal set of involved agents

This is a required question

Does your planner support private propositions in the initial state?

Yes
No

This is a required question

Does your planner support private propositions in the goal?

(That is private goals.)

Yes
No

This is a required question

Input/Output

Input of my planner can be in form of:

(a) PDDL (limited features)
(b) PDDL (limited features) with an extension defining agents
(c) PDDL (limited features) with extensions defining agents and private/public information
Other:

This is a required question

Output of my planner can be in form of:

(a) linear sequence of actions (possibly of different agents)
(b) synchronized parallel sequences of actions (each agent has an action at each time step + noop actions)
(c) partial-order plan
(d) per-agent policies
Other:

This is a required question

My planner is?

(a) satisficing (output is a sound plan)
(b) action count / plan length optimal (output is a plan with the smallest number of actions)
(c) makespan optimal (output is a shortest plan possible w.r.t. number of actions in parallel)
(d) cost optimal (output is minimal w.r.t. sum of action costs)

This is a required question

Metrics

My planner focuses on:

(a) coverage (as many as possible solved problems)
(b) minimization of planning time (solving some problems fast, but maybe not many of them)
(c) minimization of communication load among agents
(d) minimization of action count / plan length (only for satisficing planning)
(e) minimization of plan makespan (only for satisficing planning)
(f) minimization of plan cost (only for satisficing planning)
Other:

This is a required question

Technical

Please, check all answers which your planner is capable of now, and also for which you are willing to modify your planner for the competition.

(computation) In my planner:

(a) all agents run sequentially on one thread in one process (no distribution)
(b) each agent runs on own thread(s) in one process (parallel)
(c) each agent runs in its own process (distributed)

This is a required question

(computation ‒ per agent parallelism) In my planner:

(a) each agent runs on a fixed number of threads (e.g., one for computation, one for communication; or single-threaded agents)
(b) each agent can utilize variable number of threads (the computation of one agent can be parallelized)

This is a required question

(communication) In my planner:

(a) agents communicate by function/method calls (within one process)
(b) agents communicate over shared memory
(c) agents communicate by message passing over network

This is a required question

My planner runs on:

(a) *nix-like OS (Linux, OSX, ...)
(b) Windows OS
(c) needs additional SW, specify in Other. (e.g. message mediator, virtual machine, lib, etc.)
Other:

This is a required question

My planner supports running from command-line (similarly as IPC planners).

(or you are willing to modify it so)

yes
no

This is a required question

My planner is written (mostly) in programming language(s), specify.

This is a required question

The source code of my planner can be published.

yes
no

This is a required question

Additional

Additional comments

Feel free to fill in any additional comments and remarks

This is a required question

References

[1] Brafman, Ronen I., and Carmel Domshlak. "From One to Many: Planning for Loosely Coupled Multi-Agent Systems." ICAPS. 2008. [2] Ronen Brafman and Uri Zoran, Distributed Heuristic Forward Search with Interacting Actions, Proceedings of 2nd DMAP Workshop, 2014.