An AI for Diplomacy

1
An AI for Diplomacy

• Multiple players
• Deterministic
• Unrestricted Negotiation
• Projective Objective Create a Diplomat. Provide
  treatment of all aspects of Diplomacy

2
Diplomacy Overview

• Developed 1950's
• Set in WW1
• Risk-like, but drew more inspiration from Chess
• Unique

3
Diplomacy

• Game map Provinces/ Supply Centres
• Seven Players
• Units armies and fleets
• Objective, to occupy more than half of Europe

4
Diplomacy Turn Phases

• Negotiation Phase
• Order Writing Phase
• Order Resolution Phase
• (Build/Disband units Phase)

5
Diplomacy Challenges

• Size of search space
• Richness of negotiation
• Treatment of deceit
• Incomplete information

6
Diplomacy Approach

• Diplomats are neither altruistic nor malicious
• Diplomats only view winning the game outright as
  a goal

• Strategy and Tactics
• Negotiation
• Deceit

7
Strategy and Tactics

• Order Generation

Influence Maps Additional Heuristics
Plan Composition
Combinatorial Explosion Clustering
Plan Evaluation
Position Prediction Position Evaluation
8
Negotiation Space

• IOUs
• Important role in game
• Verifiability

9
Negotiation Protocol

• Negotiation Setting One time 'encounter'
• Extendable to fuller negotiation
• Encounter is a, sealed, first-price auction

Random Ordering Secrecy Private
Valuations Reserve Price IOUs as bids
10
Negotiation Strategy

• Egotistical valuations, based on strategy and
  tactics
• Auction every unit
• Auction reserve price set to loss incurred
• Bid to true valuation
• Order bids lowest to highest
• IOU valuation includes 'power' of opposing
  Diplomat

11
Deceit

• How to decide which agreements to keep/break?
• How to react when agreements broken?
• Honesty and Reliability

12
Demonstration

• Implemented using DialoX, April
• Server, Mapper, created by DAIDE

13
Findings

• Strategy and Tactics

Beat HoldBot as Turkey in 1942, Russia in
1905 Beat ConsBot 1911 Beat DAIDE AI
26/60 Eliminated 57 times vs 79 times.
14
Findings
Negotiation
Effectiveness 133 Games NN 88 of wins as Turkey
vs 76 for N Diplomat 1214 Potential
Agreements 76 Bids 46 Agreements Agreement roles
well distributed
15
Findings
Effectiveness 148 Games As Germany deceitful
Diplomat 15.8 wins, honest Diplomat 6.9 wins 6
Deceitful Diplomats 77 Games 15.2 wins achieved
with low honesty, 6.4 with high honesty 6
Honest, but aware of deceit Diplomats 71
Games Indifferent performance


16
Summary of Findings

• Plays solid strategic and tactical Diplomacy
• Negotiates effectively
• Deceives effectively
• Polarisation of effectiveness
• Useful model of a Diplomacy player.

17
Further Work

• Further trials involving present Diplomat
• Refined strategy
• Extend negotiation space
• Extend negotiation mechanism to support
  argumentation
• Allow Diplomats to vary honesty

18
Conclusion

• Achieved objective, developed a novel Diplomat
• Findings are encouraging
• Huge potential for further work
• Thanks!
• Concluding Remarks