Interview
|
by David Ploog
|
The questions are organized under four different headings: (1) questions related to Stephen Tavener's work as the developer of Ai Ai, (2) questions related to Stephen Tavener's work as a game designer, (3) general questions about Stephen Tavener as a person, and (4) interesting supplemental questions. The questions are italicized.
The Ai Ai developer
Many readers know you from the software Ai Ai, a program playing many abstract board games. When did you start on Ai Ai? How many games does it play by now?
2015, kind of. Long, long, ago, around 2012, I started a PhD on Artificial Intelligence at Imperial College. Together with Cameron Browne, we designed MoGaL (Modular Game Library), which was a general game system formed from bits of Java code glued together by a JSON-like language. Unfortunately, that wasn't a good time for the Tavener household, and I had to abandon the PhD; but a couple of years later, I dusted off the code and turned it into Ai Ai. There are still bits of MoGaL there—mostly Cameron's graphics, which are much more professional than anything I would have come up with; and the games in the Unoptimised folder still use the original MoGaL routines.
It is still in development. I have a long, long TODO list of internal features, maybe some neural net support later on.
As of April 2020, Ai Ai plays around 172 games, though it depends on how you count variants. Here is a game candidate list; here is a list of (ranked) implemented games.
What programming language do you use, and why? Would you still choose it?
Java. It's fast; maybe not quite as fast as C++/C but within about 10% for the applications I'm interested in. It's portable; the same code runs on Windows, Mac, Linux. It's also really easy to add external game and AI classes. I would probably still choose it. There are a few upcoming languages like Go that I'd consider instead, but I don't think any of them are sufficiently stable yet. Java has proved that it's here to stay.
How large is the program?
In lines of code: UI 21,702; analysis library 17,450; core libraries 26,860; search library 2,621; games: 124,093.
Of course, Ai Ai is very much a kitchen sink application, and there's a lot of code in there that isn't strictly necessary (like the fractal generator and Game of Life game end animations). Oh, and the search library currently contains 24 different AI's. 😊
As Ai Ai is a general purpose engine, it plays different games at very different levels. What kinds of subgenre does it excel at; where are its weak points?
Ai Ai excels at games with a small state space, small boards (usually), and games that are fairly opaque to players, so Ai Ai doesn't have to think hard to beat them. So, the Ai Ai AI's are AlphaBeta (min-max pruning) or UCT (Upper Confidence Trees) based; and they have different strengths and weaknesses. UCT is the default, because there is little game-specific knowledge required, but it is driven by random playouts, so anything that affects the speed (complex rules, large branching factor, long games) affects the playing strength too. Example: Gomoku on a 19x19 board. Games which have one or two good moves amongst a lot of bad moves can also cause problems; movement/capture games like Chess often fall into this category. AlphaBeta is the fallback when UCT fails. In this case, I need to hand-craft the heuristics for the game—which can make the difference between implementing a game in a few hours, and implementing a game in a few days.
Which games needed a lot of additional thought, care and code?
Almost anything by Nick Bentley (Bug, Carnivores, Blooms) is likely to cause me trouble. Chess-like games (Navia Dratp, Hermetica, Tank Chess, Cheaoss have all been problematical—I still need to revisit Hermetica and improve the heuristics. The most surprising game was Breakthrough, which was really tough—I think because humans can see the patterns and read ahead a long way.
I still can't tell you which games Ai Ai plays well—give or take a few trivial ones which are solved, I could make the AI stronger with more effort for any game you mention. People spend years making the perfect Chess engine (etc.) whereas I usually just spend a few hours; so I don't feel too bad about that!
I can tell you which is the worst game—Gess. The combinatorics are horrendous, and even with heuristics it plays very badly. I've rewritten the code from scratch once, and I'll rewrite it again at some point. [Update from Stephen just before publication: “After the second rewrite, it finally plays a decent game!" ~ Ed.]
Artificial Intelligence development has a branch called General Game Playing (GGP) which deals with programs not dedicated to a single game. Clearly, Ai Ai belongs there. GGP has many developers, some competitions and there is on-going research. Do you have connections to this scene?
Connections? A little: I did a little work for Zillions of Games back around the turn of the millennium, and I've done a little work for the Digital Ludeme Project. To a large extent, though, the Academics ignore me and I ignore them. Ai Ai probably doesn't qualify as a general game player in the strictest sense; most GGP programmes start with a language, and play games defined in that language; in Ai Ai, the language is little more than a set of JSON files used as glue to stick pieces of Java together. It can behave like a GGP if all the necessary modules are there, but more frequently I write the main game code in Java and just reference it from the JSON files for efficiency.
Why, you ask? Well, Zillions is undoubtedly the most successful GGP to date; you can implement almost any game with it, and I implemented quite a few back in the day. It does, however, have a number of missing features—connection goals and integer variables are the big ones. If you want to implement a game with either of these features you end up writing a lot of code; and Zillions will play the game very poorly as a result. I decided to go for a completely open ended system instead—all components in the system interact through a set of Java interfaces, and any game satisfying the game interface will work efficiently with any of the AI's. I can write optimized code for a single game, or an engine that will play a number of game variants with a minimum of configuration.
About the engines: I have an engine which supports several Go variants, another for Tafl variants, one for n-in-a-row games, one for the Tix family of games [see Tix in this issue], and I'm currently writing one for Chess variants.
What about connection games (Hex, Havannah, etc.) and elimination games (Draughts or Checkers or Fanorona)—these don't need specialized engines?
Hex and Y have their own engine, but it's very optimized and won't generalize very far. Similarly, there are some cute optimization tricks I use in Havannah that would cost clock cycles instead of saving them in other games. UK Draughts fits nicely in an 8x8 grid, and has an efficient bitboard implementation. International Checkers uses a 10x10 grid, and needs a different internal representation. Also, maximal capture is a significant challenge.
Tell us about Ai Ai's optimisation!
In essence, Ai Ai expects a game to provide very little information; how many moves are available at a given node in the game tree, what's the state (ongoing/won/lost/drawn, etc), apply the n-th move. There's a bit more, but most of it is optional.
Each game/engine has to provide this information, and there is a strong trade-off between generalization and speed ("premature optimisation is the root of all evil," they say, but this is one of the few cases where every clock cycle counts!). If I can generalize usefully without hurting performance, I try to do so and make an engine, but very often it's better to write very specific code for each game.
When adding games to Ai Ai, do you prefer new designs or would also you consider classics? For example, I know people who'd love to play Star, *Star and Poly-Y on Ai Ai; myself, I'd be so happy for Epaminondas or Focus.
If the game is still in copyright I usually only implement games if I can get permission from the designer/publisher, or if there is a clear signal that the game is in the public domain. Not always though: Connect 4 would be a case in point; n-in-a-row games are the drosophila of the AI community and Connect-4 has been part of Ai Ai since before I had a coherent policy on these matters.
With traditional games, it's clearly not a problem. With recent games, the author is usually willing to oblige though some publishers are reluctant to do so. There's that area in the middle, though, where someone owns the copyright, but they're not easy to contact. Ea Ea is an example—if someone can get permission from him, I'd be happy to add his games to Ai Ai. As mentioned above, the game of Y is already in.
The multitude of available games can be overwhelming. How do you deal with it? As a coder, how do you decide which games to implement?
True. The rate of new game creation has been increasing for a while. It's not just in the small world of abstract gaming, either.
There are three main reasons I implement a game:
A related question: how do you find games within Ai Ai?
I so often get feedback that it's hard to know where to start. That's the main reason for the list of ranked games.
You can also find related games through the Categories menu, and there's a search option in the File menu where you can type a keyword and find related games.
Many readers know you from the software Ai Ai, a program playing many abstract board games. When did you start on Ai Ai? How many games does it play by now?
2015, kind of. Long, long, ago, around 2012, I started a PhD on Artificial Intelligence at Imperial College. Together with Cameron Browne, we designed MoGaL (Modular Game Library), which was a general game system formed from bits of Java code glued together by a JSON-like language. Unfortunately, that wasn't a good time for the Tavener household, and I had to abandon the PhD; but a couple of years later, I dusted off the code and turned it into Ai Ai. There are still bits of MoGaL there—mostly Cameron's graphics, which are much more professional than anything I would have come up with; and the games in the Unoptimised folder still use the original MoGaL routines.
It is still in development. I have a long, long TODO list of internal features, maybe some neural net support later on.
As of April 2020, Ai Ai plays around 172 games, though it depends on how you count variants. Here is a game candidate list; here is a list of (ranked) implemented games.
What programming language do you use, and why? Would you still choose it?
Java. It's fast; maybe not quite as fast as C++/C but within about 10% for the applications I'm interested in. It's portable; the same code runs on Windows, Mac, Linux. It's also really easy to add external game and AI classes. I would probably still choose it. There are a few upcoming languages like Go that I'd consider instead, but I don't think any of them are sufficiently stable yet. Java has proved that it's here to stay.
How large is the program?
In lines of code: UI 21,702; analysis library 17,450; core libraries 26,860; search library 2,621; games: 124,093.
Of course, Ai Ai is very much a kitchen sink application, and there's a lot of code in there that isn't strictly necessary (like the fractal generator and Game of Life game end animations). Oh, and the search library currently contains 24 different AI's. 😊
As Ai Ai is a general purpose engine, it plays different games at very different levels. What kinds of subgenre does it excel at; where are its weak points?
Ai Ai excels at games with a small state space, small boards (usually), and games that are fairly opaque to players, so Ai Ai doesn't have to think hard to beat them. So, the Ai Ai AI's are AlphaBeta (min-max pruning) or UCT (Upper Confidence Trees) based; and they have different strengths and weaknesses. UCT is the default, because there is little game-specific knowledge required, but it is driven by random playouts, so anything that affects the speed (complex rules, large branching factor, long games) affects the playing strength too. Example: Gomoku on a 19x19 board. Games which have one or two good moves amongst a lot of bad moves can also cause problems; movement/capture games like Chess often fall into this category. AlphaBeta is the fallback when UCT fails. In this case, I need to hand-craft the heuristics for the game—which can make the difference between implementing a game in a few hours, and implementing a game in a few days.
Which games needed a lot of additional thought, care and code?
Almost anything by Nick Bentley (Bug, Carnivores, Blooms) is likely to cause me trouble. Chess-like games (Navia Dratp, Hermetica, Tank Chess, Cheaoss have all been problematical—I still need to revisit Hermetica and improve the heuristics. The most surprising game was Breakthrough, which was really tough—I think because humans can see the patterns and read ahead a long way.
I still can't tell you which games Ai Ai plays well—give or take a few trivial ones which are solved, I could make the AI stronger with more effort for any game you mention. People spend years making the perfect Chess engine (etc.) whereas I usually just spend a few hours; so I don't feel too bad about that!
I can tell you which is the worst game—Gess. The combinatorics are horrendous, and even with heuristics it plays very badly. I've rewritten the code from scratch once, and I'll rewrite it again at some point. [Update from Stephen just before publication: “After the second rewrite, it finally plays a decent game!" ~ Ed.]
Artificial Intelligence development has a branch called General Game Playing (GGP) which deals with programs not dedicated to a single game. Clearly, Ai Ai belongs there. GGP has many developers, some competitions and there is on-going research. Do you have connections to this scene?
Connections? A little: I did a little work for Zillions of Games back around the turn of the millennium, and I've done a little work for the Digital Ludeme Project. To a large extent, though, the Academics ignore me and I ignore them. Ai Ai probably doesn't qualify as a general game player in the strictest sense; most GGP programmes start with a language, and play games defined in that language; in Ai Ai, the language is little more than a set of JSON files used as glue to stick pieces of Java together. It can behave like a GGP if all the necessary modules are there, but more frequently I write the main game code in Java and just reference it from the JSON files for efficiency.
Why, you ask? Well, Zillions is undoubtedly the most successful GGP to date; you can implement almost any game with it, and I implemented quite a few back in the day. It does, however, have a number of missing features—connection goals and integer variables are the big ones. If you want to implement a game with either of these features you end up writing a lot of code; and Zillions will play the game very poorly as a result. I decided to go for a completely open ended system instead—all components in the system interact through a set of Java interfaces, and any game satisfying the game interface will work efficiently with any of the AI's. I can write optimized code for a single game, or an engine that will play a number of game variants with a minimum of configuration.
About the engines: I have an engine which supports several Go variants, another for Tafl variants, one for n-in-a-row games, one for the Tix family of games [see Tix in this issue], and I'm currently writing one for Chess variants.
What about connection games (Hex, Havannah, etc.) and elimination games (Draughts or Checkers or Fanorona)—these don't need specialized engines?
Hex and Y have their own engine, but it's very optimized and won't generalize very far. Similarly, there are some cute optimization tricks I use in Havannah that would cost clock cycles instead of saving them in other games. UK Draughts fits nicely in an 8x8 grid, and has an efficient bitboard implementation. International Checkers uses a 10x10 grid, and needs a different internal representation. Also, maximal capture is a significant challenge.
Tell us about Ai Ai's optimisation!
In essence, Ai Ai expects a game to provide very little information; how many moves are available at a given node in the game tree, what's the state (ongoing/won/lost/drawn, etc), apply the n-th move. There's a bit more, but most of it is optional.
Each game/engine has to provide this information, and there is a strong trade-off between generalization and speed ("premature optimisation is the root of all evil," they say, but this is one of the few cases where every clock cycle counts!). If I can generalize usefully without hurting performance, I try to do so and make an engine, but very often it's better to write very specific code for each game.
When adding games to Ai Ai, do you prefer new designs or would also you consider classics? For example, I know people who'd love to play Star, *Star and Poly-Y on Ai Ai; myself, I'd be so happy for Epaminondas or Focus.
If the game is still in copyright I usually only implement games if I can get permission from the designer/publisher, or if there is a clear signal that the game is in the public domain. Not always though: Connect 4 would be a case in point; n-in-a-row games are the drosophila of the AI community and Connect-4 has been part of Ai Ai since before I had a coherent policy on these matters.
With traditional games, it's clearly not a problem. With recent games, the author is usually willing to oblige though some publishers are reluctant to do so. There's that area in the middle, though, where someone owns the copyright, but they're not easy to contact. Ea Ea is an example—if someone can get permission from him, I'd be happy to add his games to Ai Ai. As mentioned above, the game of Y is already in.
The multitude of available games can be overwhelming. How do you deal with it? As a coder, how do you decide which games to implement?
True. The rate of new game creation has been increasing for a while. It's not just in the small world of abstract gaming, either.
There are three main reasons I implement a game:
- Curiosity; I've seen a game that I want to analyse for some reason; most recently Shobu (unfortunately, the publishers have an app in the pipeline, so I've not been given permission to release it).
- By request; I pick up games from the list of candidates when I fancy implementing a game or two. Thumbs help me prioritize but ease of implementation, how well I expect Ai Ai to be able to play, permission from the designer, and how much I like the game are also factors.
- The Combinatorial Game Awards; the current one is here. I usually aim to implement as many of the finalists as possible (designers willing), to help folks make an informed choice.
A related question: how do you find games within Ai Ai?
I so often get feedback that it's hard to know where to start. That's the main reason for the list of ranked games.
You can also find related games through the Categories menu, and there's a search option in the File menu where you can type a keyword and find related games.
The game designer
You are a prolific designer yourself. Which of your creations make you particularly proud? If a reader is willing to inspect three Tavener games, which ones do you recommend?
Crosshairs, Mutton, Web of Flies, CHEaoSS, Sleepers, Four, and Knight Line are all what I wanted them to be, and I'm proud of them all.
Crosshairs captures the feeling of aerial combat, while being a game of pure skill. There are plenty of opportunities for big turns. Sadly, it's out of print at the moment but you can still play in Ai Ai.
Sleepers seems to be my most popular game. It's not combinatorial; there's random tile draw, hidden information, and even a little bit of memory but it can be intensely psychological and there are so many fun combos you can pull off that you want to keep playing.
Knight Line has an elegant simplicity to it, both in the rules and the components. All you need is a couple of stacks of square counters, and you're ready to go. At each turn, split a stack and move the top half a knight's move away (playing area must remain connected at all times). Make a line of four to win. Knights being what they are, Chess tactics can creep in; forks, removing the guard. There are also ladders and territorial elements to the game. 😊
What do you think about new takes on old themes: Chess, Checkers, connection?
There's a great essay by T. S. Elliot, in which he says: "Immature poets imitate; mature poets steal; bad poets deface what they take, and good poets make it into something better, or at least something different." [The Sacred Wood: Essays on Poetry and Criticism, T. S. Elliot, 1920]
It's the same with game design. If a designer takes existing ideas and remixes them into something that plays differently (requires different ways of thinking, different strategies), that's great. If they make changes—no matter how large—that leave me playing the same game (or, worse, an inferior version) then why bother?
There are various qualities usually associated with modern abstracts: simplicity, depth, clarity, draw avoidance, balance. But you didn't specify theme! I am asking because your own games lean on the thematic side. Do you have an explicit reason for this?
I didn't mention theme so as not to muddy the waters; as a community, we can't agree on what an abstract strategy game is. 🙂
So... yes, most of my games have a motif or theme running through them. As you may infer from the T. S. Elliot quote, I regard game design as a creative endeavour. As such, my game designs usually begin with an idea/theme/motif that won't go away. Four, ironically, is my third attempt to create a game based on the four colour theorem (see also Chroma; the other didn't work at all and has been consigned back to the depths). Crosshairs started with the phrase "caught in the crossfire" together with a disappointing experience playing Wings of War. Knight Line started with the desire to create the simplest game I could (in terms of material). Itsy was similarly motivated by an attempt to simplify Go as far as possible.
I believe that an abstract game has a higher chance to catch on if it has an intrinsic theme [See “Stories and Themes for Boardgames” in AG18] . What do you think?
I want to make some distinctions here. When people talk about theme in the wider hobby they are often talking about the dressing (beautiful plumage). This leads us to lots of soulless Euros with pretty artwork and nice wooden components where you are converting one colour cube into another until eventually you win by having a pile of victory points. A good theme helps you play a game, and the moves build a narrative/tell a story as the game progresses.
This can be hard to do with a lot of abstracts, but as you play Chess, for instance, it's easy to imagine a battle between two armies, fought over a series of exchanges, until one side is put into a hopeless position and capitulates.
This is one reason why I used the word "motif" earlier. Four is an example where there is a motif (4-colour theorem) but no narrative. Santorini is an example where there is a sense of narrative, of a story unfurling as the game progresses. The recently kickstarted Thrive (formerly Eigenstate) has a nonsense dressing involving ponds, which really doesn't resonate with me at all.
What is your opinion on why some games survive and others perish?
The enthusiasm of the designer and/or publisher is certainly a factor. Over and above that? I really don't know. People are strange. Most recent case in point is Shobu which seems to have made the crossover from hardcore abstract gamers to the wider hobby. The driftwood-and-pebbles presentation helps, and I suspect the lack of clarity probably does so as well. There's plenty of scope to turn around an apparently bad position that maybe appeals to a wider audience.
So you'd say that drama is important for an abstract's appeal these days?
In the wider community, certainly. In a discussion on Chess a while back, a friend said "I'm not a strong player but I like Chess because I can go along to a club and find a game at my level." Most of the modern abstracts don't have clubs, so you'll likely be playing with one or two friends who are willing to play two-player strategy games. If one player wins consistently, that game isn't going to get table space. This happened with my wife and I, playing Zèrtz for instance—we started out on an equal footing, and enjoyed the game; but then something clicked and I started seeing long forcing sequences. She had no chance after that point, and we stopped playing. A good handicapping mechanism may help, but if there are other games where a handicap is not required, why not play those instead?
So, drama helps level the playing field, and gives a player at a lower skill level more chance of winning. This is also why many of the most popular games in the abstract category on BoardGameGeek tend to have random elements.
Sleepers has 61 hexes, Mutton 37 and Crosshairs uses a base 6 hex-hex board. I get the feeling you prefer smaller boards?
Personally, I think there's something very interesting about a small board, even in games like Go. It distils the game down to its essence, and allows you to learn tactics and strategies that can then be applied to a larger board.
And speaking with my AI programmer's hat on; scaling up increases both the branching factor and game length, so it really hurts the AI strength; Ai Ai likes small boards too! In human play, it depends on the game—sometimes it just makes them longer and reduces the clarity without increasing the strategy. I also like precision; finding the right move, rather than one of several that are good enough.
I'd also say that clarity increases. Oh, and you get straight to the action; some games have fairly long opening phases where the moves are pointless unless you're a strong player... and sometimes even then.
Also a factor, but even in Eurogames this is controversial: "We think that it is important that every single turn should matter. If you cannot lose the game on the first turn, then the first turn should not be played." (Joris Wiersinga, Splotter Spellen)
What's the smallest game you recommend?
Stacking is cheating, as it effectively increases the board size vertically, so the smallest good game I can think of offhand is Gyges, which is 6x6. There are very strong tactical elements (big turns), but strategic placement is important from the start once you get past the stupid blunder stage. It may also be that I really, really like tactics—I've played a lot of Chess over the years.
I recently did some analysis of games on a 3x3 board using Ai Ai, and found some surprisingly interesting games. Martian Tic Tac Toe, Traffic Lights, and Tak are all interesting, and bust the 3x3 restriction by using stacking. Tic Tac Chess is also quite challenging, (though the challenge is actually for the second player not to lose.)
Some games introduce a unique mechanism, such as Reversi's toggle capture, Hex's connection goal, the unification goal of Lines of Action, the Symple movement protocol. Each of these concepts is powerful enough to spawn successors. I'm very interested in mechanisms like these. Do you know others?
Shogi piece recycling. Action points—Conquest may be the first? Boardless boards such as Chex, Trax and many more modern mechanisms. The 12* protocol which is older than Connect 6, but that's where it was named. Dice as counters—Duell may be the first. Groups-as-organisms/perceptual binding: Carnivores, Bugs. I wonder what the first stacking game was? I can't think of any historic games with a stacking mechanism. Lately, I was impressed by Shobu, with the do something on one board, repeat on a different board. And the Sy- mechanism (grow each group, create a new one) which I shamelessly stole for Itsy.
What about some games or mechanisms you don't like?
I really dislike the chicken auction at the start of games like Unlur and Lyngk. Try not to make good moves for as long as possible? Ugh!
And cold games? Hate them!
Rithmomachy—so tedious, no wonder it was abandoned in favour of Chess!
Do you think that the quality of abstract designs has generally gone up?
Interesting question. I did a little research by looking through my ratings on BoardGameGeek, and these are my favourite games by year—not necessarily the ones I rated highly at the time, but the ones I still pull out and play when the opportunity arises.
You are a prolific designer yourself. Which of your creations make you particularly proud? If a reader is willing to inspect three Tavener games, which ones do you recommend?
Crosshairs, Mutton, Web of Flies, CHEaoSS, Sleepers, Four, and Knight Line are all what I wanted them to be, and I'm proud of them all.
Crosshairs captures the feeling of aerial combat, while being a game of pure skill. There are plenty of opportunities for big turns. Sadly, it's out of print at the moment but you can still play in Ai Ai.
Sleepers seems to be my most popular game. It's not combinatorial; there's random tile draw, hidden information, and even a little bit of memory but it can be intensely psychological and there are so many fun combos you can pull off that you want to keep playing.
Knight Line has an elegant simplicity to it, both in the rules and the components. All you need is a couple of stacks of square counters, and you're ready to go. At each turn, split a stack and move the top half a knight's move away (playing area must remain connected at all times). Make a line of four to win. Knights being what they are, Chess tactics can creep in; forks, removing the guard. There are also ladders and territorial elements to the game. 😊
What do you think about new takes on old themes: Chess, Checkers, connection?
There's a great essay by T. S. Elliot, in which he says: "Immature poets imitate; mature poets steal; bad poets deface what they take, and good poets make it into something better, or at least something different." [The Sacred Wood: Essays on Poetry and Criticism, T. S. Elliot, 1920]
It's the same with game design. If a designer takes existing ideas and remixes them into something that plays differently (requires different ways of thinking, different strategies), that's great. If they make changes—no matter how large—that leave me playing the same game (or, worse, an inferior version) then why bother?
There are various qualities usually associated with modern abstracts: simplicity, depth, clarity, draw avoidance, balance. But you didn't specify theme! I am asking because your own games lean on the thematic side. Do you have an explicit reason for this?
I didn't mention theme so as not to muddy the waters; as a community, we can't agree on what an abstract strategy game is. 🙂
So... yes, most of my games have a motif or theme running through them. As you may infer from the T. S. Elliot quote, I regard game design as a creative endeavour. As such, my game designs usually begin with an idea/theme/motif that won't go away. Four, ironically, is my third attempt to create a game based on the four colour theorem (see also Chroma; the other didn't work at all and has been consigned back to the depths). Crosshairs started with the phrase "caught in the crossfire" together with a disappointing experience playing Wings of War. Knight Line started with the desire to create the simplest game I could (in terms of material). Itsy was similarly motivated by an attempt to simplify Go as far as possible.
I believe that an abstract game has a higher chance to catch on if it has an intrinsic theme [See “Stories and Themes for Boardgames” in AG18] . What do you think?
I want to make some distinctions here. When people talk about theme in the wider hobby they are often talking about the dressing (beautiful plumage). This leads us to lots of soulless Euros with pretty artwork and nice wooden components where you are converting one colour cube into another until eventually you win by having a pile of victory points. A good theme helps you play a game, and the moves build a narrative/tell a story as the game progresses.
This can be hard to do with a lot of abstracts, but as you play Chess, for instance, it's easy to imagine a battle between two armies, fought over a series of exchanges, until one side is put into a hopeless position and capitulates.
This is one reason why I used the word "motif" earlier. Four is an example where there is a motif (4-colour theorem) but no narrative. Santorini is an example where there is a sense of narrative, of a story unfurling as the game progresses. The recently kickstarted Thrive (formerly Eigenstate) has a nonsense dressing involving ponds, which really doesn't resonate with me at all.
What is your opinion on why some games survive and others perish?
The enthusiasm of the designer and/or publisher is certainly a factor. Over and above that? I really don't know. People are strange. Most recent case in point is Shobu which seems to have made the crossover from hardcore abstract gamers to the wider hobby. The driftwood-and-pebbles presentation helps, and I suspect the lack of clarity probably does so as well. There's plenty of scope to turn around an apparently bad position that maybe appeals to a wider audience.
So you'd say that drama is important for an abstract's appeal these days?
In the wider community, certainly. In a discussion on Chess a while back, a friend said "I'm not a strong player but I like Chess because I can go along to a club and find a game at my level." Most of the modern abstracts don't have clubs, so you'll likely be playing with one or two friends who are willing to play two-player strategy games. If one player wins consistently, that game isn't going to get table space. This happened with my wife and I, playing Zèrtz for instance—we started out on an equal footing, and enjoyed the game; but then something clicked and I started seeing long forcing sequences. She had no chance after that point, and we stopped playing. A good handicapping mechanism may help, but if there are other games where a handicap is not required, why not play those instead?
So, drama helps level the playing field, and gives a player at a lower skill level more chance of winning. This is also why many of the most popular games in the abstract category on BoardGameGeek tend to have random elements.
Sleepers has 61 hexes, Mutton 37 and Crosshairs uses a base 6 hex-hex board. I get the feeling you prefer smaller boards?
Personally, I think there's something very interesting about a small board, even in games like Go. It distils the game down to its essence, and allows you to learn tactics and strategies that can then be applied to a larger board.
And speaking with my AI programmer's hat on; scaling up increases both the branching factor and game length, so it really hurts the AI strength; Ai Ai likes small boards too! In human play, it depends on the game—sometimes it just makes them longer and reduces the clarity without increasing the strategy. I also like precision; finding the right move, rather than one of several that are good enough.
I'd also say that clarity increases. Oh, and you get straight to the action; some games have fairly long opening phases where the moves are pointless unless you're a strong player... and sometimes even then.
Also a factor, but even in Eurogames this is controversial: "We think that it is important that every single turn should matter. If you cannot lose the game on the first turn, then the first turn should not be played." (Joris Wiersinga, Splotter Spellen)
What's the smallest game you recommend?
Stacking is cheating, as it effectively increases the board size vertically, so the smallest good game I can think of offhand is Gyges, which is 6x6. There are very strong tactical elements (big turns), but strategic placement is important from the start once you get past the stupid blunder stage. It may also be that I really, really like tactics—I've played a lot of Chess over the years.
I recently did some analysis of games on a 3x3 board using Ai Ai, and found some surprisingly interesting games. Martian Tic Tac Toe, Traffic Lights, and Tak are all interesting, and bust the 3x3 restriction by using stacking. Tic Tac Chess is also quite challenging, (though the challenge is actually for the second player not to lose.)
Some games introduce a unique mechanism, such as Reversi's toggle capture, Hex's connection goal, the unification goal of Lines of Action, the Symple movement protocol. Each of these concepts is powerful enough to spawn successors. I'm very interested in mechanisms like these. Do you know others?
Shogi piece recycling. Action points—Conquest may be the first? Boardless boards such as Chex, Trax and many more modern mechanisms. The 12* protocol which is older than Connect 6, but that's where it was named. Dice as counters—Duell may be the first. Groups-as-organisms/perceptual binding: Carnivores, Bugs. I wonder what the first stacking game was? I can't think of any historic games with a stacking mechanism. Lately, I was impressed by Shobu, with the do something on one board, repeat on a different board. And the Sy- mechanism (grow each group, create a new one) which I shamelessly stole for Itsy.
What about some games or mechanisms you don't like?
I really dislike the chicken auction at the start of games like Unlur and Lyngk. Try not to make good moves for as long as possible? Ugh!
And cold games? Hate them!
Rithmomachy—so tedious, no wonder it was abandoned in favour of Chess!
Do you think that the quality of abstract designs has generally gone up?
Interesting question. I did a little research by looking through my ratings on BoardGameGeek, and these are my favourite games by year—not necessarily the ones I rated highly at the time, but the ones I still pull out and play when the opportunity arises.
|
-2200 Go 1475 Chess 1883 Othello 1889 Boxes 1942 Hex 1962 Twixt 1967 Sprouts 1974 Connect 4 1975 Blockade 1977 Mentalis* 1979 Hyle/Entropy* 1980 Trax, Interplay 1981 Havannah 1982 Chex*, Mandala* 1983 Square Routes, Kamisado* 1985 Gyges, Chase 1986 Emergo, Gute Nachbarn, Plateau* 1988 Axiom 1991 Quarto 1992 Amazons, Fibonacci, Billabong 1993 Pylos, Char 1994 Ko-An, Gess, Quantum Game* 1995 Schlangennest 1996 Gipf 1998 High Kings of Tara 1999 Octi, Zèrtz, *Star, Entrapment, Croak!* 2000 Hive 2001 Dvonn, Meander, Breakthrough 2002 Fire and Ice, Farlander, Arimaa 2003 Six, Connect 6 2004 Santorini, Navia Dratp, Darter, Tasso* 2005 Khet, Hinguere, Pilare 2006 Exxit 2007 Tzaar, Tortuga, Barca, Mambo 2008 Qyshinsu: Mystery of the Way 2010 Web of Flies, Kings Plate, Crosshairs, Ayu, Morelli, Splits, Battle Sheep*, Wizard Hex 2011 Cheaoss 2012 Four, Sleepers* 2013 Cairo Corridor, Knight Line, Six Making, Schachadma 2014 Onitama 2015 Carnivores, Vault 2016 Tak, Santorini 2017 Bug 2018 Blooms, Hermetica, Trailmazer, Tank Chess, Boom & Zoom 2019 Shobu Games with an element of randomness or hidden information that are not strictly combinatorial are marked with an asterisk*; games by Stephen are italicized. |
Overall, the number of good games per decade seems to be rising, but a lot more slowly than the number of games published, I think.
The person
I know that there are real-life abstract games meetings in London. How strongly are you involved with them? How many players usually attend? What games are played?
There's an abstract games meetup group that in normal times meets once a month in London—I'm the one who kicked the meetings off, and I usually bring the lion's share of the games with me—so yes I'm strongly involved. 😊
The numbers are slightly strange—Meetup lists 177 members of the group, but the most we've had is eight players at a session. More usually, there are around four of us—so, not a big group. The joy of these games of course is that you only need two players. A bigger problem is an odd number of players. 😊
The abstract games group is very much on hold now because of COVID-19. I look forward to reviving it as soon as the world returns to normal.
Before COVID-19, we used to play in a restaurant inside Waterloo Station on a Saturday afternoon; it's a great venue in terms of being easy to reach, classier and (more importantly) quieter than your average pub meeting. It is however quite limiting—if our numbers ever grew to double digits, we'd have to find a new venue.
I try to rotate the games each time, so there's always a different selection. Games played in the past have included:
Amazons, Axiom, Barragoon, Chamelequin, Chase, Circle of Life, Darter, Dvonn, Gyges, Fibonacci, Gipf, Go, Green, Hive, Kamikaze Chess, King's Plate, Ko-An, Linear Pursuit, LOT, Murus Gallicus, Navia Dratp, NXS, Onitama, Othello, Paco Sako, Pylos, Quarto, Red, Santorini, Shobu, Sixmaking, Trax, Tasso, Trailmazer, Tzaar.
There are other abstract games groups in London; Chess, Go, and Shogi all have regular meetings that I know of, but I'm not involved in any of those.
Did you have ever have a "lifestyle game" (abstract or not)?
No, never—as I kid I was already splitting my time between Chess, Bridge, Scrabble, and Dungeons and Dragons. I played a lot of Magic: the Gathering for a while (and still play, 25 years later), but even then I was still playing a lot of other games.
What are your hobbies besides board games?
Taijiquan, painting miniatures, baking chocolate cakes, weight lifting. I also read a lot of SF, Fantasy, and detective fiction maybe 100 books per year these days. I have one cat and only the one snake but that belongs to my eldest child. I love puzzles! Particularly Slitherlink. When it comes down to it, puzzles describes almost everything I do. 😊 Not a hobby: I am autistic.
I believe that you pursue games and Ai Ai in your spare time. What is your profession?
I'm a computer programmer by inclination; sadly, I spend most of my time telling other people how to write code rather than writing it myself these days. As you say, Ai Ai is a hobby—a chance to write code, rather than watching other folks do it! Game design is quite strongly linked to Ai Ai—if I have an idea for a game design, I'll code it and test it.
Like your game design and coding, writing about games is a spare time matter for me. This has the big advantage that I'm free: I can pursue strange decisions. Do you have a similar feeling of freedom?
Indeed, yes 😊. I implement games that interest me, get to try out new AI techniques (most of which don't work), and indulge in micro-optimizations—most of which I couldn't do in a commercial setting.
You've mentioned autism. What is your experience with that?
I haven't been formally tested, but my oldest child has. There's a genetic element, and I have other traits that are strongly associated with the same gene, like prosopagnosia (face blindness). The realization puts a lot of my life into perspective, and I wish I'd figured it out earlier than I did! I think autism is very common in both computing circles and amongst abstract gamers; you need an atypical way of thinking to excel at either, as well as a strong ability to focus on a problem. Autism is a superpower!
Supplemental questions
Which games have you added in the last few months?
As of October 2020, Ai Ai plays around 228 games, though it depends on how you count variants. Here is a list of (ranked) implemented games. Every year I try to implement (designers willing) as many of the 2019 combinatorial award entries as possible. This year, these were:
I know that there are real-life abstract games meetings in London. How strongly are you involved with them? How many players usually attend? What games are played?
There's an abstract games meetup group that in normal times meets once a month in London—I'm the one who kicked the meetings off, and I usually bring the lion's share of the games with me—so yes I'm strongly involved. 😊
The numbers are slightly strange—Meetup lists 177 members of the group, but the most we've had is eight players at a session. More usually, there are around four of us—so, not a big group. The joy of these games of course is that you only need two players. A bigger problem is an odd number of players. 😊
The abstract games group is very much on hold now because of COVID-19. I look forward to reviving it as soon as the world returns to normal.
Before COVID-19, we used to play in a restaurant inside Waterloo Station on a Saturday afternoon; it's a great venue in terms of being easy to reach, classier and (more importantly) quieter than your average pub meeting. It is however quite limiting—if our numbers ever grew to double digits, we'd have to find a new venue.
I try to rotate the games each time, so there's always a different selection. Games played in the past have included:
Amazons, Axiom, Barragoon, Chamelequin, Chase, Circle of Life, Darter, Dvonn, Gyges, Fibonacci, Gipf, Go, Green, Hive, Kamikaze Chess, King's Plate, Ko-An, Linear Pursuit, LOT, Murus Gallicus, Navia Dratp, NXS, Onitama, Othello, Paco Sako, Pylos, Quarto, Red, Santorini, Shobu, Sixmaking, Trax, Tasso, Trailmazer, Tzaar.
There are other abstract games groups in London; Chess, Go, and Shogi all have regular meetings that I know of, but I'm not involved in any of those.
Did you have ever have a "lifestyle game" (abstract or not)?
No, never—as I kid I was already splitting my time between Chess, Bridge, Scrabble, and Dungeons and Dragons. I played a lot of Magic: the Gathering for a while (and still play, 25 years later), but even then I was still playing a lot of other games.
What are your hobbies besides board games?
Taijiquan, painting miniatures, baking chocolate cakes, weight lifting. I also read a lot of SF, Fantasy, and detective fiction maybe 100 books per year these days. I have one cat and only the one snake but that belongs to my eldest child. I love puzzles! Particularly Slitherlink. When it comes down to it, puzzles describes almost everything I do. 😊 Not a hobby: I am autistic.
I believe that you pursue games and Ai Ai in your spare time. What is your profession?
I'm a computer programmer by inclination; sadly, I spend most of my time telling other people how to write code rather than writing it myself these days. As you say, Ai Ai is a hobby—a chance to write code, rather than watching other folks do it! Game design is quite strongly linked to Ai Ai—if I have an idea for a game design, I'll code it and test it.
Like your game design and coding, writing about games is a spare time matter for me. This has the big advantage that I'm free: I can pursue strange decisions. Do you have a similar feeling of freedom?
Indeed, yes 😊. I implement games that interest me, get to try out new AI techniques (most of which don't work), and indulge in micro-optimizations—most of which I couldn't do in a commercial setting.
You've mentioned autism. What is your experience with that?
I haven't been formally tested, but my oldest child has. There's a genetic element, and I have other traits that are strongly associated with the same gene, like prosopagnosia (face blindness). The realization puts a lot of my life into perspective, and I wish I'd figured it out earlier than I did! I think autism is very common in both computing circles and amongst abstract gamers; you need an atypical way of thinking to excel at either, as well as a strong ability to focus on a problem. Autism is a superpower!
Supplemental questions
Which games have you added in the last few months?
As of October 2020, Ai Ai plays around 228 games, though it depends on how you count variants. Here is a list of (ranked) implemented games. Every year I try to implement (designers willing) as many of the 2019 combinatorial award entries as possible. This year, these were:
|
Linear Pursuit Iris Wunchunk Pinch Exo-Hex Poka Yoke Chess+ Mapmaker Alliances Wizard's Garden Mammalath Tabula Trivium Keil Othellito Jaleo Make Muster Benediction |
Of the finalists, only Shobu (permission refused) and Uptown (it's 3D, which I can't do proper justice to at present) are missing.
After that, some Polish researchers drew my attention to some inefficiency in my older implementations, I did a round of going back and cleaning up some older games with newer tools that I've developed. The most notable achievement there is that my Gess implementation may actually beat a human being occasionally!
Shiny new games that I've implemented are Meadow (Nick Bentley), Slyde (Mike Zapawa), and Furl (Stephen Tavener). Older games that I've implemented are Fanorona, Canadian Checkers, and International Checkers. Arimaa made a brief appearance, but there's a license so I pulled it again. I have asked the designer for permission, but haven't heard back from him.
Some small games I've implemented for my 3x3 UCT [Upper Confidence bounds for Trees] rankings are Knight Through, Hexapawn, Traffic Lights, Achi, Three Mens Morris, and Tic Tac Chess. Coming soon is ChessLine (a generalization of Tic Tac Chess with improved play)
You have used Ai Ai to set up a UCT based skill ranking for small games. Can you say something about how and why that works?
There's a recurrent topic on the BoardGameGeek forums, where players try to assess the depth of a game in terms of the number of skill levels it can support. The discussions seldom go anywhere, because depth is one of those "I know it when I see it" things that seem hard to pin down to a single metric. It occurred to me, however, that I could side-step the issue and define something concrete.
UCT is a form of AI which builds a tree from random simulations, focusing on the branches of the tree that give the best results. It has a number of properties that make it very suitable for building this kind of skill chain that other AIs lack:
This means that I can start with a very weak AI (analogous to random play) and very slowly increase the strength, marking levels as they occur.
What are the limits of this very interesting approach, for theory and for practice?
Given sufficient processing power, this process will return a consistent metric that can be applied to any game. How closely this measures the human concept of depth is a bit more problematical. I have been running some tests for very small games on my laptop for the past few weeks, focusing on very small (3x3) games which are (mostly) easily solved; so I am in a position to give some examples. Here are some 'pure' 3x3 games (those with a single piece type) I have investigated:
After that, some Polish researchers drew my attention to some inefficiency in my older implementations, I did a round of going back and cleaning up some older games with newer tools that I've developed. The most notable achievement there is that my Gess implementation may actually beat a human being occasionally!
Shiny new games that I've implemented are Meadow (Nick Bentley), Slyde (Mike Zapawa), and Furl (Stephen Tavener). Older games that I've implemented are Fanorona, Canadian Checkers, and International Checkers. Arimaa made a brief appearance, but there's a license so I pulled it again. I have asked the designer for permission, but haven't heard back from him.
Some small games I've implemented for my 3x3 UCT [Upper Confidence bounds for Trees] rankings are Knight Through, Hexapawn, Traffic Lights, Achi, Three Mens Morris, and Tic Tac Chess. Coming soon is ChessLine (a generalization of Tic Tac Chess with improved play)
You have used Ai Ai to set up a UCT based skill ranking for small games. Can you say something about how and why that works?
There's a recurrent topic on the BoardGameGeek forums, where players try to assess the depth of a game in terms of the number of skill levels it can support. The discussions seldom go anywhere, because depth is one of those "I know it when I see it" things that seem hard to pin down to a single metric. It occurred to me, however, that I could side-step the issue and define something concrete.
UCT is a form of AI which builds a tree from random simulations, focusing on the branches of the tree that give the best results. It has a number of properties that make it very suitable for building this kind of skill chain that other AIs lack:
- You can stop any time; for AlphaBeta and Minimax searches, you must complete a pass or the results are unreliable.
- Again, unlike AlphaBeta search, there is no need for heuristics which would introduce human bias.
- UCT is known to approach perfect play as the computing resources invested increase (albeit slowly); this is not necessarily true of neural nets.
This means that I can start with a very weak AI (analogous to random play) and very slowly increase the strength, marking levels as they occur.
What are the limits of this very interesting approach, for theory and for practice?
Given sufficient processing power, this process will return a consistent metric that can be applied to any game. How closely this measures the human concept of depth is a bit more problematical. I have been running some tests for very small games on my laptop for the past few weeks, focusing on very small (3x3) games which are (mostly) easily solved; so I am in a position to give some examples. Here are some 'pure' 3x3 games (those with a single piece type) I have investigated:
Game |
Levels |
Hexapawn |
4 |
Tic Tac Toe |
7 |
Breakthrough 3x3 |
7 |
Go 3x3 (7.5 komi) |
8 |
Hex 3x3 (no swap) |
8 |
Achi |
9 |
Three Men's Morris |
10 |
Tic Tac Chess |
10 |
Hex 3x3 (with swap) |
12 |
I'll just quickly point out a few highlights; first, there's a game worse than Tic Tac Toe; Hexapawn was apparently invented by Martin Gardner in the 1960's as an example in an article about AI; looking one move ahead is sufficient for the second player to play perfectly (and win).
Secondly, note that the universally maligned Tic Tac Toe actually has a whopping 7 levels of play which I'm sure most adults will find surprising. It would be interesting to work with young children and find out if these are really there and invisible to most adults or whether they are artifacts of the UCT algorithm—all games I tested have several skill levels just above completely random—all I can really say is that they are consistent.
Finally, I'd like to highlight the difference that the swap rule makes in Hex; the UCT skill levels jump from 8 to 12, suggesting that I'm measuring something useful!
Now on to the practical constraints. For these 3x3 games it typically took a few minutes to one hour to generate all skill levels. As the board size increases, however, the branching factor increases (often with the square of the board size) and the game length increases as well, so the difference between a 3x3 and 4x4 game is already huge. For more complex games with differentiated pieces and/or stacking on a 3x3 grid (e.g., Martian Tic Tac Toe, Traffic Lights) a 10-hour run wasn't sufficient to find the highest level of play, so most larger games will quickly run into processing and memory limits before we can fully analyse them; nonetheless, a short run for a large game may still give valuable information about the lower levels of play.
I should point out that this is a new and shiny tool, and I'm still refining my understanding of what it can tell us. My current findings are here but that's not the end of the story by any means.
Secondly, note that the universally maligned Tic Tac Toe actually has a whopping 7 levels of play which I'm sure most adults will find surprising. It would be interesting to work with young children and find out if these are really there and invisible to most adults or whether they are artifacts of the UCT algorithm—all games I tested have several skill levels just above completely random—all I can really say is that they are consistent.
Finally, I'd like to highlight the difference that the swap rule makes in Hex; the UCT skill levels jump from 8 to 12, suggesting that I'm measuring something useful!
Now on to the practical constraints. For these 3x3 games it typically took a few minutes to one hour to generate all skill levels. As the board size increases, however, the branching factor increases (often with the square of the board size) and the game length increases as well, so the difference between a 3x3 and 4x4 game is already huge. For more complex games with differentiated pieces and/or stacking on a 3x3 grid (e.g., Martian Tic Tac Toe, Traffic Lights) a 10-hour run wasn't sufficient to find the highest level of play, so most larger games will quickly run into processing and memory limits before we can fully analyse them; nonetheless, a short run for a large game may still give valuable information about the lower levels of play.
I should point out that this is a new and shiny tool, and I'm still refining my understanding of what it can tell us. My current findings are here but that's not the end of the story by any means.
Many thanks to both Stephen Tavener and David Ploog for this detailed and interesting interview. Readers can investigate Ai Ai here. Stephen has been a significant contributor to the world of abstract games for many years, obviously through Zillions of Games and Ai Ai, but also through his own games. I hope we can investigate Crosshairs or one of Stephen's other titles in a future issue. Stephen was a valued writer for the old series of Abstract Games, with articles on Zèrtz in AG6, AG7, AG8, and AG9. To finish, please see below three challenging problems found by Eric Silverman utilizing Ai Ai, for Symple (AG19), Lines of Action (AG1, and others), and Stephen's own game Knight Line. The solutions are here.
Symple: Black to play and win in 6 moves (noting that a player may place several stones in a single move); P=4.
|
Lines of Action: Black to play and win in 9 moves.
|
Knight Line: White to play and win in 9 moves.
The rules of Knight Line, quoted directly from the Knight Line BoardGameGeek page are as follows: "Knight Line is a minimalist connection game for two players. Each player has 20 square tiles in their colour; at the start of the game, these tiles are placed in two adjacent stacks, one of each colour. A move consists of moving part of a stack (at least one tile must be left behind) to an empty space by making a knight's move. All pieces must be connected at least diagonally throughout the game. The first player to make a line of 4, orthogonally or diagonally, wins. To offset the first player advantage, they must move only a single piece on their first move." ~ Ed.