3- ANAC Human Agent Negotiation League

EXTENDED Deadline for submission of Human Agent Negotiation League is June 12th, 2017 at 23:59 UTC-12 !

Motivation

The Human-Agent Negotiation (HAN) competition is proposed in order to further explore the strategies, nuances, and difficulties in creating realistic and efficient agents whose primary purpose is to negotiate with humans. Previous work on human-agent negotiation has revealed the importance of several features not commonly present in agent-agent negotiation, including retreatable and partial offers, emotion exchange, preference elicitation strategies, favors and ledgers behavior, and myriad other topics. To understand these features and better create agents that use them, this competition is designed to be a showcase for the newest work in the negotiating agent community.

Summary

The HAN competition will involve each author or group of authors submitting an agent that will be tested against human subjects in a study run through the University of Southern California. The subject pool will be taken from the standard populace available on Amazon’s Mechanical Turk (MTurk) service, with normal filtration done for participants who are ineligible (see Subject Selection, below).

All agents must be compliant with the IAGO (Interactive Arbitration Guide Online) framework and API, which will allow standardization of the agents and efficient running of subjects on MTurk. Agents will all be run on the same single or set of multi-issue bargaining tasks, examples of which are included below (Domain Examples).

Agents will be allowed to communicate on several channels, including a set of natural language utterances that have been pre-selected and curated by the ANAC committee. Other channels include the exchange of offers through visual cues and natural language, preference statements, and emotional displays.

IAGO API

IAGO is a platform developed by Mell and Gratch at the University of Southern California. It is intended to serve as a testbed for Human-Agent negotiation specifically. IAGO is a web-based servlet hosting system that provides data collection and recording services, a human-usable HTML5 UI, and an API for designing human-like agents.

New tutorials have been added for IAGO! Source code and an example agent may also be downloaded from the form at the main IAGO page! In addition, a full documentation of IAGO is available from the download site, which is available at http://people.ict.usc.edu/~mell/IAGO. A brief summary is included here.

All agents are capable of using the API to send Events. Events are interpreted by the UI in preset ways that allow a human user to interpret an agent’s intentions. Human users also generate Events that are passed to the agent developer to interpret as desired. Example Events include:

• SEND_MESSAGE _ sends a natural language utterance to be displayed on the chat log. Agents may send any language they wish, while human participants are restricted to sending from a preset list of utterances.
• SEND_OFFER _ sends an encoded offer for the multi-issue bargaining task wherein all items are assigned to either the human player, the agent, or an “undecided” section of the offer table. Also sends a pre-coded, descriptive message when sent from the agent to the human player.
• SEND_EXPRESSION _ sends an emoticon (either Happy, Angry, Surprised, or Sad) to the chat log, and also briefly shows the corresponding emotion on the visual avatar of the agent.
• OFFER_IN_PROGESS _ provides information that the other player is currently crafting an offer. Must be explicitly sent by the agent developer to the human player.

All Events may be sent with a delay, to allow chaining of related events (for example, an agent designer could send a message, then wait 2 seconds, then follow-up with an offer and an expression simultaneously). Flood protection will prevent messages from being sent too frequently.

Further detail may be found in the IAGO documentation.

Subject Selection and Data Treatment

Competition subject participants will be selected from the MTurk subject pool. Subjects will be adults in the US (18 years or older), and will assert that they are permanent residents of the US (this will be verified with IP address tracking). Restriction to the US will be done to reduce cross-cultural effects. Each agent will be tested against 25 participants. Participants will not be re-used or be matched against more than one agent.

Due to the fact that MTurk participants will be US-restricted and natural language statements are used in the utterance set of the competition, participants will also be asked to affirm that their first language is English.

Subject basic demographic information will be collected, and they may be asked a set of verification questions/attention checks to ensure they comprehend and are engaged in the negotiation. Subjects who fail these questions will be removed from the competition and the resulting data set.

The data set collected by the competition organizers will be released to all agent developers/researchers, as with all submitted source code. Researchers not wishing to release source code should contact the organizers directly.

Competition Winners and Evaluation

A set of prizes will be awarded to the winners of the competition in two categories.

The first category will be the High Scoring Agent category. The winner will be determined by the agent that, at the end of the bargaining time, has achieved the highest score.

The second category will be the Agent Likeability Category. The winner will be determined by the agent that, following the conclusion of the negotiation and a subsequent survey, rates highest on user feedback questions.

These questions will include questions like:

• I would use the system again in the future.
• I cannot recommend this system to others.
• I think that I would like to use this system frequently.
• I liked my negotiation partner.
• I felt like I could trust my negotiation partner.

We maintain the opportunity to examine other categories for “bonus” prizes.

Domain Examples

We present here two example domains. A domain similar to these will be selected as the official challenge to the community.

All challenges this year are multi-issue bargaining tasks, which means both the agent and the human participant will negotiate over the same set of items. Items may have differing values to each side. A “full offer” means that all items are assigned to either the agent or the human participant. A “partial offer” means that some items remain on the table and undecided. No offer is considered binding until both players accept the same full offer.

A negotiation will only end when such a full offer is accepted, or the 10-minute time limit for the negotiation has expired. Human participants will have a warning shown when there is only 1 minute remaining. Agents will have access to the current negotiation time at all points, accurate within approximately 5 seconds. In the case that time expires with no full offer, each player will take points equal to their respective Best Alternative To Negotiated Agreement (BATNA).

Note that the IAGO API allows agent designers to read the natural language descriptions of the issues at runtime (e.g., “Issue1” can be understood to be something like “Lumber” or “Luxury Cars”). However, agents will make use of domain-agnostic calls.

Domain Example 1: This challenge is a simple multi-issue bargaining task over resources between two countries. There will be four distinct resources, with five items in each category. The items will have images and descriptions identifying them as either “Oil”, “Iron”, “Foodstuffs” or “Lumber”. The human player will assign a value of 4 points to each Oil, 3 points to each Iron, 2 points to each Lumber, and 1 point to each Foodstuff. The agent player will assign a value of 4 points to each Foodstuff, 3 points to each Lumber, 2 points to each Iron, and 1 point to each Oil. Each player’s BATNA is equal to the value of a single one their highest item (4, for both the human and the agent).

Domain Example 2: This challenge is a smaller task with greater disparity toward values, and more unknowns for the agent player. The human and the agent take on roles as partners at an estate sale. There are three distinct issues. The first issue is “Luxury Cars”, and there are 6 items in this category. The second issue is “Famous Paintings”, and there are 6 items in this category as well. The final issue is “Mansion”, and there is only 1 item in this category. The agent assigns a value of 5 points to each Luxury Car, 3 points to each Famous Painting, and 8 points to the Mansion issue. The human player assigns 6 points to Luxury Cars, 2 points to Famous Paintings, and 8 points to Mansion. That results in the same total in this “mostly distributive” task. The agent and the human both have a BATNA equal to 20 points.

Note than in both domains, the human’s point values and BATNA will NOT be revealed to the agent designers prior to the competition.

Natural Language Utterances

• It is important that we both are happy with an agreement.
• I gave a little here; you give a little next time.
• We should try to split things evenly.
• We should each get our most valuable item.
• Accept this or there will be consequences.
• Your offer sucks.
• This is the last offer. Take it or leave it.
• This is the very best offer possible.
• I can’t go any lower than this.
• We should try harder to find a deal that benefits us both.
• There’s hardly any time left to negotiate!

Additional Rules

Competition participants will be given a test scenario to practice their agents with. However, to prevent hard-coding preference data into agents, a different set of utilities will be used for the actual competition.

There will be no fewer than 3 distinct issues, and no greater than 5. Each issue will have fewer than 20 items.

Issue utilities will adhere to the following rule:

where k is the total number of issues. Succinctly, this means that the total for each side would be the same if that side got every item.

It is strictly forbidden to use any technique by which an agent stores information between participants. This includes methods by which the agent may learn preferences in one game and then subsequently passes that information (through external server communication or otherwise) back to itself in future games.

All 25 participants are to be treated as fresh instances against which the same agent will be run.

Note: Participation in this competition is done in good spirit and for the furtherance of academic knowledge. Attempts to circumvent the rules described herein or described by the ANAC organizers will not be tolerated.

Reference

Mell, J., Gratch, J. (2016) "IAGO: Interactive Arbitration Guide Online", In Proceedings of the 2016 International Conference on Autonomous Agents and Multiagent Systems. International Foundation for Autonomous Agents and Multiagent Systems.

Any updates can be found on the main website for IAGO.