The games industry in 2013 was worth an estimated $93 billion USD and was predicted to
grow to $111 billion USD in 2015 [61]. This industry represents world interest in receiving
entertainment through interactive game play, and indicates the importance of entertainment
industries. A similar increase of interest in entertainment is occurring in research, as shown
by the founding of new research groups (DC Labs1) and PhD programs (IGGI2).
2.6.1
Making more human-like moves
There has been a limited amount of research into creating human-like players for a number
of games, and a number of competitions awarding a cash prize exist for success. These com-
petitions include a tournament evaluating the “naturalness” of moves in Go [77], a second
Go tournament which evaluates the “humanity” of moves [78], and a Super Mario Bros.
tournament evaluating the “human-like” natural of an players moves [125].
The original test of the “humanness” is the famous Turing test, which has been one of
the most disputed topics in artificial intelligence since its statement in 1950 [117, 140]. The
original proposal is that a artificial agent would pass the test if it could be mistaken for
a human after 5 minutes of conversing in natural language (the original specification stated
that a 70% success rate is sufficient to pass.) One of the notable weaknesses of the Turing test
however, is that some human behaviours and unintelligent, and some intelligent behaviour
is inhuman. This explains some of inhuman from strong AI players, and also leaves us in
the realisation that simulation of human play is likely significantly different from simulation
of strong play, and also significantly different from simulation of entertaining play (i.e. play
designed to entertain a human opponent.) There have been promising attempts at passing 1http://digitalcreativity.ac.uk/
the Turing test within specific gaming environments, however it is important to note that
these attempts are often using variations of the Turing test as they are normally directed at
gameplay rather than natural language [69].
Making human-like moves is a very challenging and complex concept, because defining
the human-like qualities of a move is very difficult [139]. There are currently no established
parameters to measure the human-like nature of a game move [59]. One key area in which
AI fails is in an attempt to be “believably weak”. A common route to making an AI weaker
is to reduce its time budget to make decisions, however this often has the side-effect that
it frequently misses selecting moves which would be obvious to even the weakest human
players. Effectively it fails to be believably weak, which makes it appear non-human. Rather
than simply reducing time budget to create a weaker AI, the budget should instead be used to
select a believably weak move. This further requires some categorisation of what qualifies
as a believably weak move, and how a search can be tuned to generate these moves.
2.6.2
Entertaining Play
In this context, “entertaining” refers to a combination of fun, creativity and interest in a
move. Simulating this experience in general has long been the subject of research [48].
Schmidhuber has defined a formal definition of creativity and fun [120]. Schmidhuber’s
theory is that the interest or fun associated with a given piece of information is related to
the compressibility of that information. If the information is very easy to compress, then it
appears simplistic and uninteresting, if it’s too complex, then it appears confusing and is dif-
ficult to understand or remember. Information that is in the correct range of compressibility
is both easy to remember (and possibly difficult to forget), and also challenging enough to
Model in an attempt to better model the enjoyment a player experiences when playing a
game [135]. The model contains eight different elements, each of which provide a set of
criteria for classifying the effectiveness of a specific game. While useful in categorising and
clarifying the concepts that contribute towards play experiences, it is unclear whether any
direct benefit was obtained, as the work experienced difficulty in achieving practical results,
and was limited to providing guidelines for further work.
There is also research on using heuristics to evaluate the utility of a game design [51],
however it is worth noting that these techniques all apply to classifying the player experience
due to the game design, and not specifically the opponent play.
It is also worth considering whether an agent which maintains a constant 50% win
rate against a human player is operating as an entertaining agent. If we remove all other
concerns, the it could be seen as the agent providing a competent challenge which is well
matched to the human player. However, depending on the style of play or the human player’s
understanding of the agent’s actions, it could actually be seen to be antagonising the human
player. An agent which you can never truly improve against, and always maintain an even
win rate is possibly highly annoying over a long period of time, as the human player may
see no long term improvement in their record.
2.6.3
Human-like play
Complex play from artificial agents has lead to research into how they can be made to
appear more human in their play [59, 15], largely due to the assumption that human-like
play is desirable from an opponent. As discussed in section 2.6.1, we recognise a distinction
between human-like play and optimal play, as some non-optimal play is human-like. It
is likely that in specific scenarios we desire unintelligent play from our opponents most
desirable, but rather that we would like to experience play that meets our own expectations
of intelligence and human-like behaviour.
Ikeda & Viennot recently studied Go with intent to create a more human player [76].
They had an excellent approach to dissecting the game play, starting with building an oppo-
nent model to assess not only opponent strength, but also the opponent’s expectation from
the game (i.e. what that opponent would find entertaining). They also discuss playing Gen-
tle Moveswhen facing a less skilled opponent, which is a suboptimal move chosen to avoid
a devastating defeat of an opponent. Their Gentle play variant showed promising results
when compared against a simple reduction of thinking time. We can understand a clear link
between gentle moves and human-like behaviour by considering a benevolent teacher play-
ing against a student. The teacher may make gentle moves specifically to coach the student
through the initial stages of gameplay, and thus this experience, if recognised by the human
opponent, may appear human-like, however it is dangerous to assume such recognition, as
it may be mistaken instead for unintelligent play.
More recently, Devlin et al. demonstrated combining gameplay data with MCTS to
emulate human play [52]. The research focused on collected gameplay data from the top
rated Spades game in the Google Play store (AI Factory Spades3). Previous research had
shown that the adversarial agents in AI Factory Spades behaved substantially differently to
the human players [142, 43], and work be Devlin et al. has shown a successful technique
for biasing towards human-like behaviour.
2.6.4
The importance of challenge to an entertaining experience
Level of challenge is an important part of deriving enjoyment from a game, as players gener-
While different players find different levels of challenge enjoyable due to different skill lev-
els and different tastes, it is fair to say that a complete lack of challenge or an unbeatable
challenge are undesirable scenarios for the majority of players [84, 75]. Achieving a bal-
ance when directly manipulating win rate can be challenging however, as the player can feel
like they are receiving a diminished play experience due to artificial constraints on win rates
forcing AI losses [49, 43].