[agents] The third Supply Chain Management League Compeition (Part of ANAC) in conjunction with IJCAI 2022

Yasser Mohammad yasserfarouk at gmail.com
Wed Mar 16 00:04:08 EDT 2022 

You are invited to submit your agent to the third Supply Chain Management
League (SCML) running as part of the thirteenth Automated Negotiating
Agents Competition (ANAC) in conjunction with IJCAI 2022.
Challenge

Design and build an autonomous agent that negotiates on behalf of a factory
manager situated in a supply chain management simulation. The goal of a
factory manager in SCML is to maximize its profit given its private
production capabilities by negotiating trades with other agents. A factory
manager can engage in several negotiations simultaneously, for which its
utility functions are in general interdependent. These negotiations, and
any ensuing contracts, are bilateral. Moreover, they are private to the
agents involved.
Resources

For more information about SCML including game description, tutorials,
submission instructions, etc; please refer to the competition website:
https://scml.cs.brown.edu
Important Dates

   - The official release of the competition platform (python only) *March
   15st, 2022*
   - The website opens for submission *March 15st, 2022*
   - Tournaments start *March 15th, 2022*
   - Registration on the competition website (Recommended): *April 30th,
   2022*
   - Preliminary submission deadline (REQUIRED): *May 21st, 2022*
   - Final submission deadline: *June 10th, 2022*
   - Academic Report submission deadline: *June 17th 2022*
   - Notification to finalists: *July 1st*
   - Announcement of winners: during IJCAI 2022

Changes from SCML 2021

Each agent’s score in the collusion track will be the difference in its
performance between when it controls one factory and when it controls three
factories in the same production level, averaged across multiple
simulations. This change reduces the dependence of an agent’s score in the
collusion track on its standard strategy.
Negotiation Protocol

Agents are factory managers that control factories with private
manufacturing profiles which are revealed at the start of each simulation.
Factory manager agents negotiate bilaterally with other agents to buy the
necessary inputs to their manufacturing process, and to sell the outputs.

All negotiations are carried out via the alternating offers protocol. This
protocol specifies that two negotiators take turns making offers. One agent
starts the negotiation with an opening bid, after which the other party can
take the following actions: - Accept the offer - Make a counteroffer, thus
rejecting and overriding the previous offer - Walk away, thus declaring an
end to the negotiation without having reached an agreement

This process is repeated until either an agreement is reached, or the
deadline arrives. To reach an agreement, both parties must accept the
offer. If no agreement has been reached by the deadline, the negotiation
fails.

A single simulation runs for a predefined number of steps with an overall
time limit of two hours. All negotiations are conducted for a predefined
number of rounds of the alternating offers protocol (with a predefined time
limit on each).

Factory manager agents are reset after each simulation. This means that
they cannot learn from previous simulations. They can, however, accumulate
information about agents during a simulation, as they know their
negotiating partners’ names.
Platform

Entrants to the competition should develop and submit an autonomous agent
that runs on NegMAS. NegMAS is a Python-based negotiation platform in which
you can create simulated worlds, like the SCM world, populated with agents
capable of engaging in multiple negotiations.

In NegMAS, the alternating offers protocol is implemented as a special
bilateral case of the multilateral Stacked Alternating Offers Protocol
(slightly modified to allow for a limited number of rounds).

Participants will need to install the SCML library that runs on top of
NegMAS and includes the SCML2020World and SCML2020OneShotWorld
environments. Using this library, you can run the SCM world with the same
settings as the ones employed in ANAC 2021, or with personalized settings.

SCML also includes implementations of a few sample strategies, which can
serve as guides to participating teams.

Participants can also install the SCML-Agents library which contains the
source code and reports of all agents that qualified for the ANAC 2020 SCML
track.
Submission and Live Competition

An unofficial live competition will be run this year, beginning *March 15th*.
All participants are encouraged to upload early versions of their agents to
the online submission site and are required to upload a working agent by *June
10th*.

A leaderboard will be maintained, displaying the relative performance of
all submitted agents, but no identifying information about the
participating teams will be available. This website is also where the final
versions of agents should be submitted for the official competition (at
which point identifying information will become available).

Participants must also submit the following by the deadline indicated in
the “Important Dates” table below (through the online submission site):

   - Team member names, affiliations, and contact information.
   - A single zip file containing the following (Please use the provided
   skeleton):
   - Agent source code.
   - An academic report describing the agent in PDF (The academic report
   can be updated until June 17th).

Top finalists will be required to submit a poster/video presenting their
agent for possible presentation at IJCAI. Once submitted, code and the
academic report should be considered to be in the public domain and may be
incorporated in full or part in future releases of NegMAS and/or SCML
and/or SCML-Agents or any other media. In such cases, all contributors will
receive proper attribution.

*Academic Report*: Each participating team must prepare a 2-page report
describing the novel aspects of their agent according to academic standards
(consider it an extended abstract). This report will be evaluated for
soundness, novelty, and presentation quality by the organizers of this
league, but will not affect the possibility of winning (as long as the
agent is not judged to plagiarize an existing agent). Submission of this
report is an integral part of the submission process, and agents without an
associated report will be disqualified. Additionally, agents with unusually
interesting strategies — as described by the reports and judged by the
organizing committee — may be invited to participate in the finals,
regardless of their performance during the qualification round.

At a minimum, the academic report should address the following points: -
Negotiation choices: how your agent chooses what to negotiate about, when,
and with whom - Utility function(s): how your agent calculates its utility
for the various outcomes of its negotiations - Simultaneous negotiations
coordination: how your agent coordinates its behavior among multiple
simultaneous negotiations (if it does) - Risk management: any measures your
agent takes to alleviate the risks involved in trading with other agents -
Evaluation: how your agent performs against some or all of the built-in
default strategies - Collusion (If employed): Did your agent employ a
collusive strategy? If so, what was it, and was it effective?

Moreover, finalists will be required to submit a poster presenting their
agent for possible presentation at IJCAI.
Rules of Encounter

Agents will be disqualified for violating the spirit of fair play. In
particular, the following behaviors are strictly prohibited: - Accessing
any information about the simulation or other agents that is not available
through the bulletin-board or public methods and properties of the
Agent-World-Interface. - Accessing/modifying any “private” method/member in
python (these variable names are usually indicated by an initial
underscore). - Hacking or exploiting bugs in the software. - Communicating
with the agent during the competition. - Altering the agent during the
competition. - Moreover, agents that wreak havoc on the simulator
(intentionally or otherwise) may be disqualified. Agents may be
disqualified for violating the spirit of fair play.

For more information, see the ANAC Rules. Organizers of the ANAC 2021 SCML
competition will enforce these rules. We also reserve the right to
disqualify agents under ANAC’s usual circumstances.

Finally, note that ANAC’s competition rules allow multiple entries from a
single institution, but require that each agent be developed independently.
Evaluation

There will be three separate tracks in SCML 2021. An agent’s performance in
a track will be measured by its score in that track.

In the first, the standard track, at most one instantiation of each team’s
agent will run in each simulation. Agents submitted by other teams will
also participate, but again at most one instantiation of each. Some agents
provided by the organizing committee may also control some of the factories
in a simulation. The agent’s score will be the median of the profits
accrued by all its factories in all its instantiations in all simulations.

In the second, the collusion track, multiple instantiations of the same
team’s agent will run during a single simulation. The exact number of
instantiations of each may vary across simulations, and will not be
announced in advance. In this track, it is possible for multiple
instantiations of the same agent to collude with one another to try to
corner the market, or exhibit other collusive behaviors. The agent’s score
in a given simulation will be its “consolidated financial statement,”
meaning the sum total of the profits and losses accrued by all of its
instantiations in that simulation. The agent’s score will then be the
median of its score in all simulations.

In the third, the OneShot track, an agent’s inventory and balance play a
much smaller role, as they are reset at every step of the simulation. As in
the standard track, at most one instantiation of each team’s agent will run
in each simulation. Agents submitted by other teams will also participate,
but again at most one instantiation of each. Some agents provided by the
organizing committee may also control some of the factories in a
simulation. The agent’s score will be the truncated mean of the total
profits it achieves in all steps across all simulations.

In all three tracks, ties will be broken using the agents’ mean scores.

The three tracks will be conducted in two rounds, a qualifying round and a
final round. All entrants that are not judged to break any of the SCML and
ANAC submission rules will be entered into the qualifying rounds.
Top-scoring agents in the qualifying round will then be entered in the
final round. The organizing committee maintains the right to require that
agents surpass a minimum score threshold to advance to the finals or to win
one of the prizes.

The teams that build the top-scoring agents will be notified in July, with
the final results and awards announced at IJCAI 2022 in Vienna. It is
expected that finalists will send a representative to the ANAC workshop at
IJCAI 2022, whether it is virtual or in-person, where they will be given
the opportunity to give a brief presentation describing their agent. Three
awards will be announced at IJCAI 2022 (with associated monetary rewards)
corresponding to the three tracks (standard, collusion, and OneShot).

The organizing committee will determine the number of simulations needed in
each round to ensure a fair comparison among all submitted agents. All
participating agents that achieve scores that are not statistically
different from the winners’ will be inducted into the SCM league’s hall of
fame.

The latest version of the agent submitted before the competition deadline
will be used in the SCM league unless participants opt-out of the official
competition.
Questions and Answers

Please check our FAQ. You can post your questions there (preferable), or
address any concerns you prefer to remain private to Yasser Mohammad.
Organizing Committee

   - Yasser Mohammad, NEC, NEC-AIST, RIKEN & Assiut University (main
   contact)
   - Katsuhide Fujita, Tokyo University of Agriculture and Technology &
   NEC-AIST AI Collaborative Research Laboratory
   - Amy Greenwald, Brown University
   - Mark Klein, MIT
   - Satoshi Morinaga, NEC-AIST AI Collaborative Research Laboratory
   - Shinji Nakadai, NEC-AIST AI Collaborative Research Laboratory

Sponsors

   - NEC-AIST AI Collaborative Research Laboratory

*The organizing committee would like to thank Brown University for hosting
the online submission website at https://scml.cs.brown.edu
<https://scml.cs.brown.edu>.*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cs.umbc.edu/pipermail/agents/attachments/20220315/4236ade7/attachment-0001.html>

More information about the agents mailing list