elling of knowledge inspired by the Theory of Mind
(ToM). The game platform makes it easy to evaluate
strategies and compare different parametric settings.
The unique mechanics of RBC contribute to a
profound strategic depth that significantly surpasses
traditional chess. The addition of the ‘sensing’ ac-
tion, the incomplete knowledge of the opponent’s
pieces, and the uncertainty introduced by blind cap-
tures result in a complex set of possible states of the
chess board associated with the uncertainty of knowl-
edge that players must navigate. For human players,
this complexity not only requires strategic thinking
but also introduces an element of psychological war-
fare, as players must try to anticipate their opponents’
actions and decisions based on limited information.
This game is not easy for computers either. For a
comparison, classical chess has approximately 10
43
possible states, while RBC has 10
139
possible states,
which indicates that the game can be 10
96
times more
complex (Markowitz et al., 2019).
There has been a substantial amount of research
focused on perfect information games, particularly
chess, which is well-documented in the literature (Apt
and Simon, 2021; Silver et al., 2017). However, the
study of imperfect information games, such as Recon-
naissance Blind Chess, remains relatively unexplored.
The will to fill the gap in our understanding forms
the basis of the following research questions: RQ1.
How can better knowledge modeling benefit the es-
timation of an opponent’s knowledge, which results
in enhanced performance? RQ2. What are the most
effective sensing strategies to diminish the game’s in-
herent uncertainty? Finally, with the understanding
of uncertainty in knowledge acquisition, our third re-
search question is RQ3: What is the most efficient
moving strategy?
Our contributions are the following: (1) We
present a comprehensive analysis of the game, dis-
cussing potential ways to model player knowledge
and manage the expansive state space, amidst preva-
lent uncertainties. (2) We investigate various sens-
ing and moving strategies and evaluate their respec-
tive impacts on agent performance. (3) Based on
our analytical results, we choose the best strategies
for our agent, Scorca
3
, and publish it as an open
source project
4
. The agent ranks second on the global
leaderboard
5
at the time of submission of the article
3
The performance and game records can be found here:
https://rbc.jhuapl.edu/users/48973
4
The code is available on GitHub at https://github.c
om/Robinbux/Scorca with DOI 10.5281/zenodo.10412786.
All the data and supplementary material are on Zenodo with
DOI 10.5281/zenodo.10412840.
5
https://rbc.jhuapl.edu/
(23/10/2023).
The paper is organized as follows. Section 2 pro-
vides recent research on the game. Section 3 explains
knowledge modelling in the RBC game. Section 4
and 5 discuss different sensing and moving strategies,
respectively. Since the rationale for the design of our
Scorca agent is explained in Section 3, 4, and 5 along
with their analysis, we include only a summary of
the strategies chosen with references to the sections
in which the corresponding decisions were made in
Section 6. The evaluation is included in 7, followed
by the discussion in Section 8 and conclusion and fu-
ture work in Section 9.
2 RELATED WORK
A rich landscape of RBC agents have emerged, each
with its uniquestrategy and design during the NeurIPS
tournaments held in 2019 (Gardner et al., ), 2021
(Perrotta et al., 2019), and 2022 (Gardner et al.,
2022). The strategies, algorithms, and methodologies
adopted by these agents have influenced our research.
These agents provide references and benchmarks to
our approach. The following paragraphs give a sum-
mary of these agents and their respective mechanisms
based on the description in the Tournament reports.
StrangeFish
6
maintains an exhaustive set of pos-
sible board states, expanding and filtering based on
game events and private observations. It chooses
sensing actions to maximize the expected impact on
its next move selection. Its moving strategy evaluates
options across the set of boards using a weighted com-
bination of best-case, worst-case, and average out-
come scores calculated with Stockfish
7
(a widely used
chess engine) and RBC-specific heuristics. This ap-
proach of tracking uncertainty while leveraging chess
knowledge allowed StrangeFish to win the inaugural
NeurIPS RBC tournament in 2019 (Gardner et al., ).
StrangeFish2
8
, the successor to StrangeFish, re-
tains its predecessor’s framework for tracking board
states but adds several enhancements. It chooses sens-
ing actions by estimating outcome probabilities for
hypothetical moves after each possible observation.
This results in selecting the sense with the greatest
expected value based on potential move outcomes.
StrangeFish2 also improves efficiency through paral-
lelization and improved search algorithms.
The Fianchetto agent
9
(Taufeeque et al., 2022) is
6
https://rbc.jhuapl.edu/users/713 with code at https:
//github.com/ginop/reconchess-strangefish.
7
https://stockfishchess.org/
8
https://rbc.jhuapl.edu/users/1987
9
https://rbc.jhuapl.edu/users/12368
Knowledge Modelling, Strategy Designing, and Agent Engineering for Reconnaissance Blind Chess
211