Up to $250 for Feedback on Welfare Diplomacy Variant for AI Research
Posted: Tue Jul 18, 2023 12:01 am
TLDR: Awarding up to $250 in prizes for the best feedback on a general-sum version of Diplomacy we intend to use for cooperative AI research.
Hello! We're AI researchers from Stanford, Cambridge, Mila, and the Center on Long-Term Risk, and we're interested in AI research based on Diplomacy. We're designing a new variant called Welfare Diplomacy which allows you to trade off your military capabilities for the welfare of your power (more details below). This is interesting from the perspective of research intended to make AIs better at cooperating because it turns Diplomacy from a zero-sum game (compete for the fixed number of supply centers) to a general-sum game in which there's more opportunity for cooperation (by cooperating, you can create more welfare).
We're reaching out to ask for feedback on our variant from the community. We aren't particularly skilled at Diplomacy, so we thought it would be good to get input from more experienced players. We think an interesting general-sum version of Diplomacy would be a big win for cooperative AI research, and this would help us develop one faster. It could also help us deploy an interesting, human-playable version sooner. We are awarding USD $500 total in prizes ($250 for first place, $150 for second, and $100 for third) to the best feedback submissions as judged by us.
Submission Rules
To submit feedback, please read about the variant below, then send an email to [email protected] with the subject line "Welfare Diplomacy Feedback." We are particularly interested in answers to the following questions:
Welfare Diplomacy Rules
Summary: You can build fewer units than your current supply center count, and the difference each year adds to your power's Welfare Points (WP). Scoring is based on accumulated WP at the end of the game, meaning powers have incentives to trade off military conquest with making peace to mutually benefit their nations.
Hello! We're AI researchers from Stanford, Cambridge, Mila, and the Center on Long-Term Risk, and we're interested in AI research based on Diplomacy. We're designing a new variant called Welfare Diplomacy which allows you to trade off your military capabilities for the welfare of your power (more details below). This is interesting from the perspective of research intended to make AIs better at cooperating because it turns Diplomacy from a zero-sum game (compete for the fixed number of supply centers) to a general-sum game in which there's more opportunity for cooperation (by cooperating, you can create more welfare).
We're reaching out to ask for feedback on our variant from the community. We aren't particularly skilled at Diplomacy, so we thought it would be good to get input from more experienced players. We think an interesting general-sum version of Diplomacy would be a big win for cooperative AI research, and this would help us develop one faster. It could also help us deploy an interesting, human-playable version sooner. We are awarding USD $500 total in prizes ($250 for first place, $150 for second, and $100 for third) to the best feedback submissions as judged by us.
Submission Rules
To submit feedback, please read about the variant below, then send an email to [email protected] with the subject line "Welfare Diplomacy Feedback." We are particularly interested in answers to the following questions:
- What are your overall thoughts about Welfare Diplomacy?
- What strategies do you expect skilled Diplomacy players to try when starting to play this variant?
- What strategies do you expect skilled Diplomacy players to eventually adopt after lots of play with this variant?
- How would these rules change the ways you negotiate with the other players in a game?
- How likely is it that all 7 players negotiate an agreement early in the game and never deviate? What are specific agreements (in terms of supply centers assigned to each player, demilitarization schedules, etc) that seem likely to you?
- How likely is it that optimal play always results in a particular set of countries allying to take over the others?
- How likely is it that these rules lead to boring or degenerate outcomes?
- What are the implications of different max turn numbers?
- How balanced are these rules towards attackers or defenders, and what would you change to improve the balance?
- In which situations would players choose disarmament or not? What other situations or changes to the rules might make this more or less likely?
- What do you think of our possible further variations? Should we adopt any of them, and do you have other ideas to consider?
- Anything else you think we should know?
Welfare Diplomacy Rules
Summary: You can build fewer units than your current supply center count, and the difference each year adds to your power's Welfare Points (WP). Scoring is based on accumulated WP at the end of the game, meaning powers have incentives to trade off military conquest with making peace to mutually benefit their nations.
- In the build phase, you can freely disband or build any number of units (but not both building and disbanding), so long as your total unit count is less than or equal to your supply center count.
- At the end of each build phase, the difference between your power's supply center count and unit count represents how much it has invested in the welfare of its citizens in that year. Your power accumulates Welfare Points (WP) equal to this difference. WP continually add up each year—you can never lose them as they represent the past experiences of your power's citizens.
- At the end of the game, the winner is not the power with the greatest supply center count (this is very different from Classic Diplomacy). Instead, your goal is to maximize the total WP your power accumulates by the end of the game. You're not trying to get the most WP, you're trying to maximize your own WP, so it's very unlike typical games in this respect.
- This is a general-sum and (we think) more cooperative variant of Diplomacy: Rather than competing to slice up a fixed "pie," players who cooperate well can actually create more of the "pie."
- You can more explicitly make peaceful commitments to other players: Instead of just agreeing to a DMZ, two allies actually have incentives to disband their units along a demilitarized border.
- This variant possibly favors attackers: If many nations have agreed to peace and some subset secretly militarize, they have a year head-start to move against the others uncontested.
- We will need to tweak some aspects like the max number of turns in a game to balance the tradeoff between early expansion and late-game peace.
- Build Anywhere: Allow building units in all owned supply centers, not just homes, to allow for more rapid militarization.
- Overmilitarize: You can build more units than you have supply centers, but then you will lose WP according to the difference between the two.
- Frequent Building: Build phases happen after every turn, not just after every year, so if some coalition defects and rapidly militarizes, there's only 1 turn ahead of others building to catch up, not 2.
- Progressive WP Weighting: Each year, you gain WP equal to the supply-unit difference multiplied by the number of years since 1900, so 1 WP per difference in 1901, 8 per difference in 1908, etc.
- Economics Points: WP are instead Economics Points (EP) and can be sent to other powers (either whenever just like chat messages, or submitted like orders and revealed at the end of each turn).
- Other ideas from you?
- We will be judging submissions holistically and awarding prizes according to how much insight they provide into how the game is likely to be played, backed by strong arguments and evidence. Reports from play-testing, or walkthroughs of plausible sequences of play, would make especially strong submissions.
- We'll award $250 for first place, $150 for second place, and $100 for third place ($500 total) in US dollars.
- We've cross-posted this on a few different Diplomacy message boards, but we will only award 3 prizes.
- If you have a winning submission, we'll contact you via the same email you submitted to arrange payment.
- Payments will go through Wise. We can only send prizes to countries that Wise can send to. We will make all reasonable efforts to make payment, but there may be rare cases in which payment is not possible due to legal requirements (such as payments to countries or individuals under international sanctions or conflicts of interest with the Center on Long-Term Risk's board members).
- We will be posting the 3 winning solutions online for others to see afterwards.
- We will ask each contestant who suggested a change or consideration that we incorporated if they want to be included in the acknowledgments section of a research paper we will possibly produce.