What happens when a robot learns to play Diplomacy? Listen to jmo and peterwiggin (pw) from webDip on the most transformative online Dip development in many years. This is a game changer folks! Plus talk on WDC in Marseilles, DipCon in Seattle, the 2nd face-to-face game in Brisbane and panhandle closure. Phew... this is a packed episode!
Intro
The guys channel Guns’n’Roses before discussing their drinks and venue (0 mins 10 secs)
They discuss their previous recording with jmo and his post-interview discussion with the guys about bringing artificial intelligence (AI) to webDip. They give the context about the academic paper and the interview with jmo and peterwiggin (pw) - the techincal brains trust supporting the implementation before diving in (2 mins 30 secs)
Interview with jmo and peterwiggin about AI and Diplomacy
Amby apologises up front as the interview they recorded with jmo and pw a week ago didn’t get recorded (4 mins 40 secs)
They get straight into what the webDip crew have been doing with the University of Montreal to develop bots that can play Diplomacy at a pretty high level on the classic map. Jmo explains how it all works at a high level and how you can play games with bots (5 mins 15 secs)
He discusses the processes for drawing (8 mins 50 secs)
Conceding has also been brought into webDip from vDip, and ratings won’t be affected. He discusses how if you don’t get enough real life players you can fill it with bots. Jmo discusses it going live on “this coming Friday” which is the day this podcast episode is released (9 mins 45 secs)
Amby asks about the dynamics for how the bot thinks, drawing from 150,000 games as part of its intelligence (11 mins 30 secs)
Pw talks about the machine learning process, neural networks, training the bots and reinforcement learning (13 mins)
Kaner confirms his understanding of how the bots open and pw discusses. They then discuss the bots’ “memory” and throwing support to the bots (21 mins 45 secs)
Kaner and Amby discuss their joint test game against the bots (Kaner was France, Amby was England) and reflect how quickly the bots started wiping them out. Amby asks about the extent the bots think about strategy and planning for the future (26 mins)
Kaner (France) & Amby’s (England) bot test game where it was up to in the interview
Amby asks about how the bots draw from previous game data sets to make its decisions (32 mins 30 secs)
Amby goes into more detail around the bots’ thinking and determining its own logic versus mathematical analysis of data (35 mins 15 secs)
Kaner asks about the bots’ names and how human players are likely to respond to AI players (37 mins 20 secs)
Amby asks about whether the bots recall its previous games, understanding what works and takes on a personality, or does it effectively reboot everytime (41 mins 10 secs)
Jmo expands and talks about how bots can be set up to fill games in webDip where a player goes into civil disorder (CD’s) (43 mins 30 secs)
Kaner asks about whether the bots and their code will be available on other Diplomacy sites (48 mins)
Amby mentions how bot games are unranked, but asks whether there will be future tournaments/league tables for human players against the bots (50 mins 30 secs)
Amby talks about his own solo game against the bots and how he’s stalemated the bots. He asks based on what’s transpired whether that would allow for a successful draw (52 mins 45 secs)
Amby’s solo game (Russia) which he stalemated before taking up Kaner’s suggestion
Kaner asks whether things would change if Amby used his Silesia unit to do something different (57 mins 15 secs)
Amby as reflects on how has got used to the Interactive Map Order Interface at vDip and found webDip’s drop down menus not as intuitive. He asks whether there’s plans to move that functionality across (59 mins)
Amby asks whether other variants - like Ancient Med - have a sufficiently large dataset to draw on to bring across to bot games (1 hr 01 mins 30 secs)
Kaner ask
