5.3 Experimental Methodology
6.2.9 Deck Creation
As a final experiment, we attempted to create a selection of novel decklists by providing a
small set of cards and allowing the algorithms above to generate the decks. In each case, a
single Identity card was provided, along with three other cards in the correct faction which
are commonly played in tournament play. The generated decks were then shown to a group
of experienced players, who were asked to criticise the decklists. All of the generated decks
are listed in Appendix B.
In each case, the experienced players were content that the decks were reasonable for
play in a casual or semi-competitive environment, but may struggle in a highly competitive
tournament environment. They agreed that there were no completely inappropriate inclu-
sions, however there were a few questionable choices, and in some cases it appeared that
some card combinations had been left unfinished in the decks. When asked to provide mod-
ifications to the decks to improve their quality, the players each suggest 1-3 card exchanges,
stating that those changes would likely make the decks worthy of play in a more competitive
6.3
Summary & Discussion
Our attempts to adequately predict the number of duplicate cards within a deck have been
some what effective, but there is still work to be done here, as our best prediction is based on
our heuristic knowledge of the specific card, rather than knowledge of the card in context.
Successfully adding contextual heuristic knowledge into this process will surely lead to
more accurate prediction.
Given that experienced players found our decks to be at least worthy of consideration
for play in a tournament, there is definite potential in taking this technology forward, and
potentially creating tools for the Netrunner community to generate and validate their decks.
With a little adaptation, this can also be taken into other similar games.
As discussed earlier in this chapter, knowledge of the content of an opponent’s deck
represents a potentially powerful strategic knowledge which can be exploited to significant
advantage. As such, we would expect an agent which can successfully predict the content
of an opponent’s deck to be significantly stronger than an agent which was unable to do so.
It is also worth noting however, that a misprediction could cause a significant reduction in
agent strength.
6.3.1
Contributions
Our principle research contribution from our work on Rule Association Mining is a substan-
tial improvement in deck prediction from the default apriori algorithm. It can be seen that
our modifications to the Apriori technique provide a significant improvement to prediction
of decks in Netrunner, showing a maximum improvement of ∼13% between the default
apriori algorithm (a1) and our optimal modified algorithm (a7).
ously, the ability to predict opponent decks can be a powerful strategic advantage, but this
can also be leveraged in other ways. An oppositional agent that attempts to predict the hu-
man player’s deck rather than “cheating” and looking is likely to feel more fair (and more
interesting) to the human player, assuming this is correctly measured. This contribution
goes further than merely deck prediction, as these techniques can also be used as a form of
deckbuilding aide, to help new players build decks and guide them as to what might be sen-
sible inclusions in their first decks. Given that deck building is a challenging task for human
players, any help provided by an machine learning agent would likely provide significant
learning for new players and assistance for experienced players in the process of building a
deck.
Similar principles could also be used in other computer games such as MOBAs (Multi-
player Online Battle Arena) and RTS (Real-time Strategy) games. Both these archetypes of
games use highly customisable and configurable Build Orders which dictate the sequence in
which actions should be taken for optimal performance. Determining build order could be
Conclusions & Further Work
In this thesis, we have investigated the application of specific artificial intelligence tech-
niques to creating differing playstyle in games. Our principle body of work has been the
modification of MCTS in order to affect play style, and thus create entertaining or interest-
ing differences in play between a modified agent and a unmodified agent. In order to do so,
we have considered a variety of techniques including heuristic pruning and the modification
of the action selection mechanism. This work is motivated in large part by the needs of
the commercial games industry and particularly those of my sponsoring company, Stainless
Games. Commercial games need entertaining AI opponents, not merely strong ones, and
this is an area ripe for research, with little work done to date.
Our work in heuristic pruning and action selection mechanisms (see chapter 5) showed
that we could create different behaviour in an artificial agent without diminishing play
strength. If our approximation of entertainment through complexity holds, then we have
also shown that we can make agent play more entertaining to humans through these modifi-
cations, which has real implications for game AI.
more efficiently, and thus can make better, more interesting decisions (whether the metric
for those decisions be play strength or complexity of play.)
Below we present and summarise our research contribution, and consider their general-
isation. Then we outline a plan for future work based on some of the topics explored in our
work here. It is not our plan to execute all these plans, but we leave them here as a indi-
cation to others who may want to continue this work. Work already under way is some of
that detailed in section 7.3.4, as the application of our techniques to other games has already
provided interesting results.