=Paper=
{{Paper
|id=Vol-2862/paper9
|storemode=property
|title=TAG: A Tabletop Games Framework
|pdfUrl=https://ceur-ws.org/Vol-2862/paper9.pdf
|volume=Vol-2862
|authors=Raluca D. Gaina,Martin Balla,Alexander Dockhorn,Raúl Montoliu,Diego Perez-Liebana
|dblpUrl=https://dblp.org/rec/conf/aiide/GainaBDML20
}}
==TAG: A Tabletop Games Framework==
https://ceur-ws.org/Vol-2862/paper9.pdf
TAG: A Tabletop Games Framework
Raluca D. Gaina* , Martin Balla* , Alexander Dockhorn* , Raúl Montoliu† , Diego Perez-Liebana*
*
School of Electronic Engineering and Computer Science, Queen Mary University of London, UK
†
Insitute of New Imaging Technologies, Jaume I University, Castellon, Spain
Abstract In this work, we introduce the Tabletop Games (TAG)
Tabletop games come in a variety of forms, including board framework, which is a collection of tabletop games, agents,
games, card games, and dice games. In recent years, their and development tools meant to provide a common platform
complexity has considerably increased, with many compo- for AI research. This work is mainly motivated by three fac-
nents, rules that change dynamically through the game, diverse tors: i) the characteristics of modern tabletop games (multi-
player roles, and a series of control parameters that influence player, partially observable, large action spaces, competition,
a game’s balance. As such, they also encompass novel and and collaboration, etc.) provide an interesting challenge to
intricate challenges for Artificial Intelligence methods, yet re- AI research. These modern games provide many characteris-
search largely focuses on classical board games such as chess tics not implemented in existing frameworks, e.g. changing
and Go. We introduce in this work the Tabletop Games (TAG) player roles or varying forms of cooperative and competitive
framework, which promotes research into general AI in mod-
game-play; ii) TAG presents tabletop games from a GGP per-
ern tabletop games, facilitating the implementation of new
games and AI players, while providing analytics to capture the spective, by providing a common API for games and playing
complexities of the challenges proposed. We include prelimi- agents; and iii) we aim to provide a framework that can in-
nary results with sample AI players, showing some moderate corporate different games under a common platform, making
success, with plenty of room for improvement, and discuss it possible for the research community to implement their
further developments and new research directions. own games and AI players to expand TAG’s collection. De
Araujo et al. (2019) highlighted the need for such a frame-
1 Introduction work in a survey that described the different schemes and
data structures used in the literature of digital board games.
In the last few decades, tabletop games have gone through a
In order to provide the necessary flexibility to support the
‘Renaissance‘, gaining more popularity than ever: thousands
great variety of existing tabletop games, TAG requires the
of them are published each year and welcomed by an ex-
user to implement games via a programming language (Java),
panding audience of gamers (Engelstein and Shalev 2019).
instead of using a game description language as other general
Many of these games are typically designed with richer me-
frameworks do. TAG provides many customisable compo-
chanics and rules, and less focus on chance-based elements.
nents to simplify the development of additional games not
Modern tabletop games are very diverse and complex in gen-
currently in the framework. It is able to handle partial observ-
eral, which can provide various challenges, such as unique
ability, providing simple means of adding custom observation
game state representations, partial observability on various
schemes, game-state heuristics, and agent statistics, as well
levels, actions outside of a player’s turn, cooperation, and
as supporting the development of graphical user interfaces
competition in the same game, etc. Therefore, they have a
for human play with computer players.
very different and complex set of mechanisms that would
While we are working on extending the framework with
require a lot of effort to develop using previous approaches.
additional game mechanics, games and agents, the current
AI Research in board games has mostly been, with some
version is publicly available1 . The contributions of this paper
exceptions, focused on traditional board games, either in
are two-fold: firstly, we present the framework, its structure,
isolation (chess, Othello, Go, etc.) or as part of general
games and AI players implemented (see Section 3). Secondly,
game playing (GGP) frameworks, such as GGP (Genesereth,
we provide a discussion on a baseline experimentation (Sec-
Love, and Pell 2005), Ludii (Piette et al. 2019) and Open-
tion 4), aimed at illustrating insights into the games imple-
Spiel (Lanctot and others 2019). While these frameworks also
mented, their features, and performance of vanilla AI players.
allow the definition of additional games, they are limited to
In Section 5 we discuss challenges and opportunities included
common mechanics or require extensive development effort.
for this framework and we conclude with final remarks on
Copyright c 2020 for this paper by its authors. Use permitted under future developments in Section 6.
Creative Commons License Attribution 4.0 International (CC BY
1
4.0). https://github.com/GAIGResearch/TabletopGames
2 Related Work Spronck 2009) and Risk (Gibson, Desai, and Zhao 2010),
AI research and board games have been closely related since and Rolling Horizon Evolutionary Algorithms (RHEA) in
the beginnings of the field, with game-playing agents for Tic- Splendor (Bravi et al. 2019), all showing a great improve-
Tac-Toe, Checkers and chess (Campbell and others 2002), and ment in performance. Other games have been more recently
the more recent breakthroughs in Go (Silver et al. 2017). One highlighted as important challenges for AI players due to
of the most well-known contests, the General Game Playing their strategic complexity (Pandemic (Chacón and Eger 2019;
(GGP) competition (Genesereth, Love, and Pell 2005), fea- Sfikas and Liapis 2020)) or very large action spaces (Blood
tured classical board games written in the Game Description bowl (Justesen et al. 2019)).
Language and promoted research into generic game-players. Research has not only focused on game-playing AI, how-
(Piette et al. 2019) later introduced the “ludemic” general ever. Ticket to Ride (de Mesentier Silva et al. 2017) was
game system Ludii, which builds upon ideas from GGP. Ludii used as an example for employing AI players for play-testing
defines games as structures of ludemes, high-level, easily un- games, characterising their features based on different play-
derstandable game concepts, which allows for concise and styles and finding possible bugs or gaps in the ruleset. Further,
human-understandable game descriptions, as well as easy the use of Procedural Content Generation for such games is
implementation of new games. The current main focus of highlighted by (Guzdial et al. 2020); given the rule com-
the Ludii project is on classical and ancient board games. plexities and the multitude of components in modern table-
Kowalski et al. presented in (Kowalski et al. 2019) a new top games, AI methods can provide a more efficient way of
GGP language, called Regular Boardgames (RBG), with a searching the possibility space for interesting variations.
similar focus as Ludii, but which describes games as regular The Tabletop Games (TAG) framework introduced in this
expressions instead, for increased efficiency. paper brings together all of these different research directions
We consider direct code implementations to be more ac- and provides a common ground for the use of AI algorithms
cessible, faster to execute and easier to work with in many in a variety of tabletop games, removing the effort of creating
cases. More similarly to our framework in this regard, Open- different frameworks for different purposes and simplifying
Spiel (Lanctot and others 2019) provides a collection of the overall development process. As far as we know, TAG is
games and algorithms written in C++ and exposed to Python. the first framework that allows the development of multiple
Most of their games are still traditional board games, with games and AI players under a common API for complex
some exceptions, such as the inclusion of Hanabi. Differ- modern tabletop games.
ently, TAG shifts the development effort onto the framework,
rather than the games, by making a wide range of compo- 3 The Framework
nents, rules, actions, etc. available to users. Our system allows
TAG was designed to capture most of the complexity that
for fast prototyping of new games and immediate extraction
modern tabletop games provide, with a few games imple-
of insights and features of interest in each game through
mented already and more in progress.
readily-available game and AI-facilitated analysis. We fur-
ther support many research directions, from simulation-based
game-playing to parameter optimisation of games and artifi-
3.1 Concepts
cial players. Our framework includes handy definitions for various
Here, we focus on more types of games, including board, concepts and components common across many tabletop
card, dice and role-playing games, among others, often games (Engelstein and Shalev 2019).
grouped under the tabletop games umbrella term. Tabletop We define an action as an independent unit of game logic
Simulator (Henry 2015) is an example of software facili- that modifies a given game state towards a specific effect (e.g.
tating implementation of tabletop games components in a player draws a card; player moves their pawn). These actions
physics-based simulated environment; this allows players are executed by the game players and are subject to certain
to interact with the games as they would in the real world, rules: units of game logic, part of a hierarchical structure (a
but the many games implemented lack support for automatic game flow graph). Rules dictate how a given game state is
rule execution, nor does the software facilitate AI research modified and control the flow through the game graph (for
as targeted with TAG. However, tabletop games research has instance, checking the end of game conditions and the turn
been gaining popularity in recent years. Research in game- order). This turn order defines which player is due to play
playing agents for card games is common in competitive at each time, possibly handling player reactions forced by
(Poker (Moravčı́k and others 2017) and Bridge (Cazenave actions or rules. At a higher level, games can be structured
and Ventos 2019)) and cooperative (Hannabi (Bard and oth- in phases, which are time frames where specific rules apply
ers 2020)) games. and/or different actions are available for the players.
Asymmetric player roles is one feature often encoun- All tabletop games use components (game objects shar-
tered in modern tabletop games, and these have been stud- ing certain properties), whose state is modified by actions
ied in games such as The Resistance (Serrino et al. 2019) and rules during the game. TAG includes several predefined
and Ultimate Werewolf (Eger and Martens 2019). This is components to ease the development of new games, such
just another complexity added in modern tabletop games, as tokens (a game piece of a particular type), dice (with N
yet more lead to the need for intricate strategic planning. sides), cards (with text, images or numbers), counters (with
To this extent, Monte Carlo Tree Search (MCTS) methods a numerical value), grid and graph boards. Components can
have been tried in Settlers of Catan (Szita, Chaslot, and also be grouped into collections: an area groups components
lenges for AI players. A few games are currently in active
development (Descent (Fantasy Flight Publishing, Inc. 2012),
Carcassonne (Wrede 2000) and Settlers of Catan (Teuber
1995)), and many more are in the project’s backlog, including
games from other frameworks to allow for easy comparison
(see Section 2).
All games implemented can be found in the games pack-
age, each registered in the games.GameType class; this
class allows specifying properties for each game, to allow for
automatic listing for experiments (e.g. a list of all games with
Figure 1: GUI for Love Letter, red line shows current player. the “cooperative” tag). We highlight next some particularities
of the games currently implemented in the framework.
Tic-Tac-Toe 2 players alternate placing their symbol in a
in a map structure in order to provide access to them using
N × N grid until one player completes a line, column or
their unique IDs, while a deck is an ordered collection with
diagonal and wins the game; if all cells in the grid get filled
specific interactions available (e.g. shuffle, draw, etc.). Both
up without a winner, the game is a draw.
areas and decks are considered components themselves.
This is the simplest game included in the framework, meant
3.2 Structure to be used as a quick reference for the minimum requirements
to get a game up and running. Its implementation makes use
The TAG framework brings together all of the concepts and of mostly default concepts and components, but it implements
components described previously and allows quick imple- a scoring heuristic and a custom GUI for an easier interaction
mentation and prototyping of new games. To this end, a given the specific game mechanics.
flexible API is provided for all functionality needed to define
a game, with multiple abstract classes that can be extended Love Letter (Kanai 2012) 2 to 4 players start the game
for specific implementations. The framework provides some with one card each, representing a character, a value and a
generic functionality, from ready-made components, rules, unique effect. A second card is drawn at the start of each
actions, turn orders and game phases, to a fully functional turn, one of which must be played afterwards. After the last
game loop and a prototyping GUI. The GUI allows users to card of the deck is drawn, the player with the highest valued
start interacting with the games as soon as they have the two card wins the current round. A player wins the game after
main classes required set up: a Game State (GS) class, and a winning 5 rounds. Love Letter features partial observability,
Forward Model (FM) class. asymmetric and changing player roles and a point system
GS is a container class, including all variables and game over several rounds. Figure 1 shows an example game state.
components which would allow one to describe one specific
moment in time. It defines access methods in the game state Uno (Robbins 1971) The game consists of coloured cards
to retrieve existing game components, make custom and par- with actions or numbers. Numbered cards can only be played
tially observable copies of the state, and define an evaluation in case either the colour or the number matches the newest
function that can be used by the playing agents. The FM card on the discard pile. Action cards let 2 to 10 players
encompasses the logic of the game: performs the game setup, draw additional cards, choose the next colour to be played or
defines what actions players can take in a particular game reverse the turn order. A player wins after gaining a number
state, applies the effect of player actions and any other game of points over several rounds (computed as the sum of all
rules applicable, uses a turn order to decide which player is other players’ card values). Uno features stochasticity, partial
due to play next (or may wait for all players to return an ac- observability and a dynamically changing turn order.
tion before processing for simultaneous-actions games), and
checks for any end of game conditions. The FM is available Virus! (Cabrero and others 2015) 2 to 6 players have a
to AI players for game simulations. body each that consists of four organs, which can be: in-
For each game, users can further implement specific ac- fected (by an opponent playing a virus card), vaccinated (by
tions, rules, turn orders, game parameters (for easy modifica- a medicine card), immunised (by 2 medicine cards) or de-
tion of game mechanics), a GUI and provision of game data. stroyed (by opponents playing 2 consecutive virus cards). The
The last is useful when the game requires large amounts of winner is the first player who forms a healthy and complete
data such as tile patterns, cards and board node connections, body. Virus! features stochasticity and partial observability,
and it is provided via JSON files. A full guide on using the with the draw pile and opponents’ cards being hidden.
framework and implementing new games is available in the Exploding Kittens (Inman and others 2015) 2 to 5 play-
wiki provided with the code and in (Gaina et al. 2020). ers try to avoid drawing an exploding kitten card while col-
lecting other useful cards. Each card gives a player access to
3.3 Games unique actions to modify the game-state, e.g. selecting the
There are currently 7 games implemented in the framework, player taking a turn next and shuffling the deck. This game
varying from very simple test games (Tic-Tac-Toe) to strategy features stochasticity, partial observability and a dynamic turn
games (Pandemic (Leacock 2008)), as well as diverse chal- order with out-of-turn actions: in contrast to previous games,
Exploding Kittens keeps an action stack so that players have IStateHeuristic interface. Custom heuristics take the
the chance to react to cards played. current state as input and return a scalar number representing
the value of that state without any other restrictions. Agents
Colt Express (Raimbault 2014) 2 to 6 players control a
can be given a heuristic function on initialisation and then in-
bandit each, with a unique special ability. Their goal is to
stead of using the reward directly from the game they process
collect the most money while traversing the two-level com-
every state they visit using the provided heuristic.
partments in a train and avoiding the sheriff (a non-playing
character moved by players and round card events). The game Human Players Two types of human interaction are avail-
consists of several rounds, each with a planning (players play able, both of which interrupt the game loop to wait for hu-
action cards) and an execution (cards are executed in the man actions on their turn. Console allows human play using
same order) phase. This processing scheme forces players the console. It outputs the game state and available actions
to adapt their strategy according to all the moves already in the console and the player inputs the index of the ac-
played, in an interesting case of partial observability and non- tion they choose to play. GUI allows human play with a
determinism: the opponents’ type of action may be known Graphical User Interface, which is game-specific. It uses
(sometimes completely hidden in a round), but not how it an ActionController object to register player action
will be executed. Additionally, the overall strategy should be requests, which are then executed in the game.
adapted to a bandit’s unique abilities.
Pandemic (Leacock 2008) Pandemic is a cooperative Random The simplest automatic player chooses random
board game for 2 to 4 players. The board represents a world actions out of those available on its turn.
map, with major cities connected by a graph. Four diseases One Step Look Ahead (OSLA) A greedy exhaustive
break out and the objective of the players is to cure them all. search algorithm, it evaluates all actions from a given game
Diseases keep spreading after each player’s turn, sometimes state and picks that which leads to the highest valued state.
leading to outbreaks. Each player is assigned a unique role
with special abilities and is given cards that can be used for Rolling Horizon Evolutionary Algorithm (RHEA)
travelling between cities, building research stations or cur- RHEA (Perez-Liebana et al. 2013) evolves a sequence of
ing diseases. Additionally, they have access to special event L = 10 actions over several generations, choosing the first
cards, which can be played anytime (also out-of-turn). All action of the best sequence found to play in the game. The
players lose if they run out of cards in the draw deck, if too algorithm is randomly initialised with a sequence of actions.
many outbreaks occur or if the diseases spread too much. At each generation it creates a mutation of the current best
Pandemic features partial observability with face-down decks solution, keeping the best solution of the two. This process
of cards and asymmetric player roles. It employs a reaction repeats until the given budget is exhausted.
system to handle event cards and is the only game currently Given the variable action spaces and that actions available
using the graph-based rule system. are highly dependent on the current game state, the mutation
operator chooses a gene in the individual (i.e. position in the
3.4 AI Players action sequence) and changes all actions from that point until
All implemented players follow a simple interface, only re- the end of the individual to new random valid actions. The
quiring one method to be implemented: getAction. This game’s forward model is therefore used in both mutation (to
receives a game state object reduced to the specific player’s advance the game state given the last action, in order to find
observation of the current state of the game. How this re- the available actions for the given position in the sequence)
duced game state is built is game-dependent, usually ran- and evaluation (game states reached through the sequence of
domising unknown information. This method expects an actions are evaluated using the game’s heuristic, added up
action to be returned out of those available and is called for a discounted total with discount factor γ = 0.9, and this
whenever it is the player’s turn and they have more than 1 total becomes the fitness of the individual). It is important to
action available (i.e. the player actually has a decision to note that RHEA evolves only its own actions and opponents
make). If no decision is required, the agent can choose to are given a random model (with intermediate states after
still receive and process the information on the game state opponent actions ignored in fitness evaluations).
(in the registerUpdatedObservation function) but
an action is not requested. They may also choose to imple- Monte Carlo Tree Search (MCTS) MCTS (Browne and
ment the initializePlayer and finalizePlayer others 2012) incrementally builds an asymmetric game tree
functions which are called at the beginning and end of the balanced towards to most promising parts of the game state
game, respectively. Each player has a player ID assigned by space. It uses multiple iterations of four steps: first, it navi-
the game engine, and they receive the forward model of the gates through the tree, using a tree policy, until reaching a
game currently being played. The FM can then be used to ad- node which is not yet fully expanded; next, it adds a new ran-
vance game states given actions, compute actions available, dom child of this node to the tree; it then performs a Monte
or reset a game to its initial state. The rest of this section Carlo rollout from the new child (randomly sampling actions
defines the sample players implemented in the framework. until the end of the game or a predetermined depth L = 10);
These agents use the game’s score to evaluate game states (as the state reached at the end of the rollout is evaluated with
implemented on the game side), but their heuristic functions a heuristic function, and this score is backed up through all
may be swapped with a different object implementing the the nodes visited during the iteration. The process is repeated
4.2 Baseline AI Player Performance
We tested the performance of the sample agents on each of
the implemented games. For Tic-Tac-Toe, we report the win-
rate per agent when playing 100 times against every possible
opponent. Since Pandemic is a cooperative game, we report
results from games played with a team of 4 instances of the
same agent, e.g. 4 MCTS players. For the remaining games,
we report the average win rate per agent when playing in a 4-
player match against one instance of all other agents. All but
the random agent are using state evaluation functions that are
provided with each game. For both search-based algorithms,
Figure 2: Action space size in Uno with all player number
we use a budget of 4000 calls to the FM.next() function.
versions; 1000 runs per version played by random players.
The average win-rate per games is shown in Table 1. Our
results indicate the MCTS agent to be the best, achieving the
until the budget is exhausted, and the most visited child of highest average win rate in 5 out of 6 competitive games, with
the root is chosen as the action to play. an overall win rate of 45.5% in these games - thus clearly
The version implemented in the framework is closed-loop: dominating the other agents. While RHEA also outperforms
it stores game states in each of the nodes. Further, the rollout random in most of the games (with the exception of Uno), it
step was removed after initial experiments showing an in- still falls behind the OSLA agent. This could be due to the
creased performance without it; therefore, the forward model large uncertainty built up in its rigid sequences of actions (as
of the game is only used when expanding a leaf node. The opposed to the flexible game trees built by MCTS) in these
resulting node is immediately evaluated using the heuristic games with partial observability and stochasticity, where
and its value is backed up through the tree. a greedy approach appears to be preferable. Further, some
games tested are simple enough that a greedy approach works
4 Discussion best and are highly advantaged by the heuristics provided by
each game (e.g. Tic-Tac-Toe).
This section gives an overview of game analytics which can Additionally, we note that no agent is able to win in the
be extracted from all games, as well as some preliminary cooperative game Pandemic, as they are unable to perform
results for the sample AI players described in Section 3.4. the multi-faceted long-term planning required to avoid all of
the loose conditions and win the game. Further analysis, such
4.1 Game Analysis
as the distance from winning game states for each AI team,
All games in the framework can be analysed to illustrate the could show interesting insights into these agents’ capabilities.
challenge they provide for AI players, with the following Moreover, we observe Uno and Colt Express as cases where
metrics currently readily available: the performance between all players, including random, is
Action space size: the number of actions available for a very close (22 − 26% in Uno and 19 − 29% in Colt Express).
player on their turn (e.g. Figure 2). Branching factor: the This highlights the difficulty of the types of problems pro-
number of distinct game states reached through player actions posed, as well as the importance of the heuristic chosen for a
from a given state. State size: the number of components game, as some features of a game state may prove deceiving.
in a state. Hidden information: the percentage of compo- However, the statistical forward planning methods de-
nents hidden from players on their turn. Game speed: the scribed here (MCTS and RHEA) benefit from ample litera-
execution speed of 4 key functions (in number of calls per ture with a large parameter space each, which could be tuned
second): setup, next, available action computation and state for increased performance.
copy. Game length: measured as the number of decisions
taken by AI players, the total number of game loop iterations
(or ticks), the number of rounds in the game and the number 5 Challenges and Opportunities
of actions per turn for a player. Reward sparsity: granularity The presented framework opens up several directions of re-
of the heuristic functions provided by the game, measured by search and proposes a variety of challenges for AI players,
min, max and standard deviation of rewards seen by players. be it search/planning or learning algorithms. Its main focus
When looking at the games currently implemented, the is to promote research into General Game AI that is able
first thing to note is that all games are very fast to execute: to play many tabletop games at, or surpassing, human level.
most games can execute over 1 million calls per second to the Relatedly, the agents should be able to handle both com-
(usually) most expensive functions (next and copy). The petitive (most common testbeds in literature), cooperative
games vary in length, with only 7.61 ticks for the simplest and even mixed games. For instance, a future planned de-
game, Tic-Tac-Toe, but 540.78 for Uno. We further see varia- velopment is the inclusion of the game Betrayal at House
tions in the state size, with Pandemic showing most complex, on the Hill (Glassco and others 2004), in which the play-
while Uno includes the most hidden information. Love Letter ers start off playing cooperatively to later split into teams
shows its strategic complexity through the higher branching mid-way through the game, from which point on they are
factor (10.78), while Exploding Kittens boasts one of the competing instead with newly given team win conditions and
largest spread of rewards. rules. Most tabletop games include some degree of hidden
Table 1: AI player performance, 100 game repetitions. Highest win rate in bold. Tic-Tac-Toe played in a round-robin tournament.
Pandemic uses 4 instances of the same agent. All others played in their 4-player variants, with 1 instance of each agent.
Tic-Tac-Toe Love Letter Uno Virus! Expl. Kittens Colt Express Pandemic Total
Random 0.12 0.00 0.26 0.01 0.05 0.19 0.00 0.09
OSLA 0.45 0.24 0.26 0.27 0.37 0.29 0.00 0.27
RHEA 0.44 0.32 0.22 0.3 0.21 0.28 0.00 0.25
MCTS 0.98 0.44 0.26 0.42 0.37 0.26 0.00 0.39
information (e.g. face-down decks of cards) and many more what was previously revealed to a player. Instead, the AI play-
players compared to traditional video-game AI testbeds, in- ers should learn to memorise relevant information and build
troducing higher levels of uncertainty. However, such games belief systems, as humans would in a real-world context - a
often make use of similar mechanics, even if in different very interesting direction of research encouraged by TAG.
forms: thus knowledge transfer would be a fruitful area Lastly, the framework includes the possibility for games
to explore, so that AI players can pick up new game rules to define their states in terms of either vector observations
more easily based on previous experiences, similar to how (IVectorObservation), which enables learning algo-
humans approach the problem. Some tabletop games fur- rithms to be easily integrated with the framework; or feature-
ther feature changing rules (e.g. Fluxx (Looney and Looney based observations (IFeatureRepresentation),
1997)) which would require highly adaptive AI players, able which allows for more complex algorithms which can
to handle changes in the game engine itself, not only the perform a search in the feature space of a game, rather than
game state. Many others rely on large amounts of content and the usual game state space approached.
components, for which the process of creating new content
or modifying the current one for balance, improved synergies
etc. could be improved with the help of Procedural Con-
tent Generation methods (e.g. cards for the game Magic the
6 Conclusion
Gathering (Garfield 1993) were previously generated in a
mixed-initiative method by (Summerville and Mateas 2016)). This paper introduces the new Tabletop Games (TAG) frame-
Specific types of games can also be targeted by research, work, which aims to promote research into general Artifi-
an option highlighted by TAG’s categorisation and labelling cial Intelligence with features for easy implementation and
of games and their mechanics. Thus AI players could learn bridge-building between tabletop games and artificial players,
to specialise in games using certain mechanics or in areas with some examples already included. We further analyse the
not yet explored, such as Role-Playing or Campaign games games in the framework, showing a wide variety of action
(i.e. games played over several linked and progressive ses- spaces, information available to the AI agents, duration etc.,
sions). These games often feature asymmetric player roles, as well as tasks and challenges introduced. The AI player
with a special one highlighted (the dungeon master) whose performance analysis shows Monte Carlo Tree Search to
aim is to control the enemies in the game in order to not dominate all other sample agents in the framework, with sim-
necessarily win, but give the rest of the players the best expe- ple greedy methods being surprisingly competitive in some
rience possible and the right level of challenge. Strategy AI of the games. Overall, however, the problems proposed are
research could see important applications in this domain, as far from being solved.
many tabletop games include elements of strategic planning. The framework opens up and facilitates many directions
Role-playing games focused more on the story created by of research, and yet many more developments are possible
players (e.g. Dungeons and Dragons) rather than combat and planned. More measurements for both games and AI
mechanics (e.g. Gloomhaven) would also be a very engaging players can be added, to paint a more complete picture of
and difficult to approach topic for AI players, where Natural the challenges the players would face, as well as the current
Language Processing research could take an important role. state of available methods for such games: skill depth, overall
The framework enables research into parameter optimi- state space size, stochasticity, size of search trees, player role
sation: all parameter classes for games, AI players or heuris- asymmetry - all would give a much more in-depth view of
tics can implement the ITunableParameters interface; the framework as a whole, aiding in future developments of
parameters can then be automatically randomised, or more both tabletop games and general AI players.
intelligently tuned by any optimisation algorithm. This al- Further, we aim to facilitate interfacing external games
lows for quick and easy exploration of various instances of with our framework in order to gain the full benefits of the in-
a problem, a potential increase in AI player performance, or depth analysis and interaction with the implemented players
adaptation of AI player behaviour to user preference. without the need to re-implement everything from scratch:
We have mentioned previously that the games imple- this would open up the framework to many more already
mented offer reduced observations of the game state to the existing games, and also increase the number and complexity
AI players, based on what they can currently observe. These of environments the AI players can be exposed to, improving
hidden information states (usually) do not keep a history of their quality as well.
Acknowledgments Artificial Intelligence and Interactive Digital Entertainment
This work was partly funded by the EPSRC CDT in Intel- Conference.
ligent Games and Game Intelligence (IGGI) EP/L015846/1 Glassco, B., et al. 2004. Betrayal at House on the Hill.
and EPSRC research grant EP/T008962/1. Avalon Hill Games, Inc.
Guzdial, M.; Acharya, D.; Kreminski, M.; Cook, M.; Elad-
References hari, M.; Liapis, A.; and Sullivan, A. 2020. Tabletop Role-
Bard, N., et al. 2020. The Hanabi Challenge: A New Frontier playing Games as Procedural Content Generators.
for AI Research. Artificial Intelligence 280:103216. Henry, J. 2015. Tabletop Simulator. Berserk Games.
Bravi, I.; Perez-Liebana, D.; Lucas, S. M.; and Liu, J. Inman, M., et al. 2015. Exploding Kittens. Ad Magic, Inc.
2019. Rinascimento: Optimising Statistical Forward Plan- Justesen, N.; Uth, L. M.; Jakobsen, C.; Moore, P. D.; To-
ning Agents for Playing Splendor. In 2019 IEEE Conference gelius, J.; and Risi, S. 2019. Blood Bowl: A New Board
on Games (CoG), 1–8. IEEE. Game Challenge And Competition For AI. In 2019 IEEE
Browne, C. B., et al. 2012. A Survey of Monte Carlo Tree Conference on Games (CoG), 1–8. IEEE.
Search Methods. IEEE Transactions on Computational Intel- Kanai, S. 2012. Love Letter. Alderac Entertainment Group.
ligence and AI in games 4(1):1–43.
Kowalski, J.; Mika, M.; Sutowicz, J.; and Szykuła, M. 2019.
Cabrero, D., et al. 2015. Virus! El Dragón Azul. Regular Boardgames. In Proceedings of the AAAI Conference
Campbell, M., et al. 2002. Deep Blue. Artificial Intelligence on Artificial Intelligence, volume 33, 1699–1706.
134(1-2):57–83. Lanctot, M., et al. 2019. OpenSpiel: A Framework for
Cazenave, T., and Ventos, V. 2019. The alpha-mu Reinforcement Learning in Games. CoRR abs/1908.09453.
Search Algorithm for the Game of Bridge. arXiv preprint Leacock, M. 2008. Pandemic. Z-Man Games, Inc.
arXiv:1911.07960. Looney, A., and Looney, K. 1997. Fluxx. Looney Labs.
Chacón, P. S., and Eger, M. 2019. Pandemic as a Challenge Moravčı́k, M., et al. 2017. Deepstack: Expert-Level Ar-
for Human-AI Cooperation. In Proceedings of the AIIDE tificial Intelligence in Heads-Up No-Limit Poker. Science
workshop on Experimental AI in Games. 356(6337):508–513.
de Araujo, L. J. P.; Charikova, M.; Sales, J. E.; Smirnov, V.; Perez-Liebana, D.; Samothrakis, S.; Lucas, S. M.; and Rolf-
and Thapaliya, A. 2019. Towards a Game-Independent shagen, P. 2013. Rolling Horizon Evolution versus Tree
Model and Data-Structures in Digital Board Games: an Search for Navigation in Single-Player Real-Time Games. In
Overview of the State-of-the-Art. In Proceedings of the Proceedings of the Genetic and Evolutionary Computation
14th International Conference on the Foundations of Digital Conference (GECCO), 351–358.
Games, 1–8.
Piette, E.; Soemers, D. J.; Stephenson, M.; Sironi, C. F.;
de Mesentier Silva, F.; Lee, S.; Togelius, J.; and Nealen, Winands, M. H.; and Browne, C. 2019. Ludii–The Ludemic
A. 2017. AI-Based Playtesting of Contemporary Board General Game System. arXiv preprint arXiv:1905.05013.
Games. In Proceedings of the International Conference on
the Foundations of Digital Games - FDG’17. ACM Press. Raimbault, C. 2014. Colt Express. Ludonaute.
Eger, M., and Martens, C. 2019. A Study of AI Agent Robbins, M. 1971. Uno. AMIGO.
Commitment in One Night Ultimate Werewolf with Human Serrino, J.; Kleiman-Weiner, M.; Parkes, D. C.; and Tenen-
Players. In Proceedings of the AAAI Conference on Artifi- baum, J. 2019. Finding Friend and Foe in Multi-Agent
cial Intelligence and Interactive Digital Entertainment, vol- Games. In Advances in Neural Information Processing Sys-
ume 15, 139–145. tems, 1251–1261.
Engelstein, G., and Shalev, I. 2019. Building Blocks of Sfikas, K., and Liapis, A. 2020. Collaborative Agent Game-
Tabletop Game Design: An Encyclopedia of Mechanisms. play in the Pandemic Board Game. In International Confer-
CRC Press LLC. ence on the Foundations of Digital Games (FDG).
Fantasy Flight Publishing, Inc. 2012. Descent: Journeys in Silver, D.; Schrittwieser, J.; Simonyan, K.; Antonoglou, I.;
the Dark 2nd Edition. Diamond Comic Distributors. Huang, A.; Guez, A.; Hubert, T.; Baker, L.; Lai, M.; Bolton,
A.; et al. 2017. Mastering the Game of Go without Human
Gaina, R. D.; Balla, M.; Dockhorn, A.; Montoliu, R.; and
Knowledge. nature 550(7676):354–359.
Perez-Liebana, D. 2020. Design and Implementation
of TAG: a Tabletop Games Framework. arXiv preprint Summerville, A. J., and Mateas, M. 2016. Mystical Tutor:
arXiv:2009.12065. A Magic: The Gathering Design Assistant Via Denoising
Sequence-To-Sequence Learning. In Twelfth artificial intelli-
Garfield, R. 1993. Magic: The Gathering. Wizards of the
gence and interactive digital entertainment conference.
Coast.
Szita, I.; Chaslot, G.; and Spronck, P. 2009. Monte-Carlo
Genesereth, M.; Love, N.; and Pell, B. 2005. General Game
Tree Search in Settlers of Catan. In Advances in Computer
Playing: Overview of the AAAI Competition. AI magazine
Games, 21–32. Springer.
26(2):62–62.
Teuber, K. 1995. The Settlers of Catan. Mayfair Games.
Gibson, R.; Desai, N.; and Zhao, R. 2010. An automated tech-
nique for drafting territories in the board game Risk. In Sixth Wrede, K.-J. 2000. Carcassonne. Hans im Glück.