Our dominance in some sectors of the games market is not to be taken for granted. Year after year our Chess Free was by far the top ranked and downloaded Chess program on Android, with some 80 million downloads. However for the first time we have seen that our #1 slot from some 200 chess titles has been challenged. A significant part of this appears to be from rival intelligent marketing, but we also realised that our popular chess was not necessarily fully accessible for complete beginners as it offered all play options right away. Once started, the tutorial system is fine, but some users might be discouraged by the complexity of the match settings screen. Also our program has no built in incentives to progress but left it to the user to move on to higher levels. We re-worked this and Chess is now gaining, still #2, but now 4 places behind our top rival, instead of 12 (in "board games"). We have much more to come so are fairly confident that our #1 status will be renewed!
As ever we dropped into Droidcon London. The highlight for me were the excellent
talks on TensorFlow, which links to the Deepmind comments below. This is easily
accessible via TensorFlowLite, which can be accessed via mobile developers.
The number of talks in this area highlights the increased importance of neural
nets in commercial work.
Spades again has taken front seat on AI development. An issue is that the same core source has progressed through multiple iterations over a period of 15 years, with 2 significant architectural shifts on the way. In consequence the game engine has some legacy game architecture not tailor-made for the current AI philosophy. We are now well connected with the beta testing community and a key part of this relationship is our capacity to do deep analysis of positions offered that appear to show defective play. These tests can take up to 2 hours to run, but provide a reliable insight into the play offered. For most of it the primary effort with Spades has been to progressively add more and more good tools for analysis. The need for this has been particularly vital for Spades as the analysis uses ISMCTS "Information set Monte Carlo tree search", which delivers a kind of "fuzzy" game tree, mostly populated with illegal variations for any one deal. This is much harder to work with than (say) conventional minimax chess, where with simple tools it is possible to exactly determine why a move was played, and fix it if needed.
The particular areas addressed in Spades in this iteration have been to improve move selection in the tree rollout through plausibility. The latter skews preference to more likely plays. The tools to analyse this provide commands in the test console to step through games to test plausibility ordering. The output is displayed in easy to absorb bar charts, rather than raw numbers. Part of the legacy update was to make sure that plausibility was exactly aligned with non-standard rules. Previously these were only approximately correct.
The article this issue examines our inference system in Spades.
We had our regular Xmas end-of-year party, this time celebrating the 15th year of the company running. Curiously the Google Console does not report total downloads any more, so not so easy to report our global downloads total. In consequence our cake was not emblazoned with our downloads total. Curiously the public Google Play store references these values, but does not think the developer should have them anymore. These can be extracted via downloaded spreadsheets but only a month at a time, so totalling all our downloads would need summing from 9 x 12 x 32 = 3456 spreadsheets! Under the old system we had 150m downloads.
Although this is still very much under-wraps we have an on-going project that expects to make use of Google Deepmind's TPU resource for creating a deep-learning game engine. This is early days, but if this comes to fruition then we expect one of our game engines may make a quantum jump into modern high performance AI.