This has been a long period with no newsletter periodical. As indicated in the header there have been new additions to family to distract us but also the massively disruptive pandemic. As early as February we were being pressed for meetings at GDC San Francisco and we then indicated that we imagined GDC would likely be cancelled. That suggestion was met with incredulity as people could not imagine that was even being considered. The idea was "crazy". It has taken a long time for people to accept the huge ramifications of this and even late April, friends were booking holidays in Spain on the basis that "this will all have blown over by May".
I think the message has now got through and happy to see people now following government advice and doing their best to take reasonable protective measures. Of course we all have families so the heavy demands of home schooling has taken its toll on time and energy, and consequently AI Factory's progress has been in the slow lane for a while.
One of the pandemic casualties: The unimaginable idea that GDC might not happen.
Despite everything, we are good and just now have completed a substantial volume of work on our Spades app. This program was started some 17 years ago and has been through many iterations. In this process it has morphed to change methods but still with some ancient game architecture in there. In consequence its iterative structure is complex. It is also likely our last game engine based on prescriptive analysis. Nowadays game AI is being taken over by deep learning techniques. Our engine actually mixes Monte Carlo (which forms part of deep learning AI) and heuristic knowledge. This blend is discussed in the previous article Mixing the Immiscible - MCTS and evaluation. It seems to be successful, allowing easy blending of radically different methods. The price paid though is that, without learning, all knowledge is derived from hand crafted input, which means there will be holes. Our intelligent beta report processing though is an effective means of finding these holes.
In addition we did also re-structure our Spades sources to work with automated tuning. This work is in place but not 100% in a state to yet make it effective. What it has done is expose that tuning complex linear evaluation code is vulnerable to defects in that code that cannot be overcome by tuning. However an intermediate goal here is that this tuning step can spot areas of linear evaluation that make no net contribution to performance and that can be weeded out, providing higher speed and deeper search. However (again) a linked issue is that Monte Carlo has a generic defect that it will tend to randomly choose between moves where all lead to the same outcome. In consequence it will then appear to have played crazy moves from time to time. The problem here is that prescriptive code is in there to compensate for this tendency, but automated tuning cannot detect this so will steer the programmer to remove code that does not compromise performance but will compromise style.
One of the good stories for us was the upgrade to our Chess. Google Play announced that netplay was no longer to be supported so we had to drop things and create an alternative using Firebase. This seemed a pointless pain at the start of the process but happily the new netplay is both faster and offers more flexibility. We expect to add proper game clocks soon and the capacity to have restricted on-line chat with pre-defined questions and responses. The latter will allow the on-line play to feel more personalised as you will have some capacity to communicate with your opponent.
As the pandemic rolled out we noticed a large volume of increased exploitative scams and a secondary casualty was our Facebook page for our Chess app. This was hacked by Cambodian hackers and sadly we have been unable to regain control of it, so apologies to our 73 thousand followers. We have a new protected page. A component in our failure to recover stems from the pandemic where Facebook do not have the bandwidth to help users suffering from such hacks.
Finally we have been experimenting with new ad networks and considering outsourcing our ad network control. Progress here has been muddied by the intervening pandemic and also we have found intermediate things we needed to try before this. We now also use Unity and Adcolony.
We wish all our partners out there well in these difficult times. Unlike for most companies, the enforced lockdowns have been less of a culture shock for us as we each work in home offices. It seems people still want to play our games and ad networks still want to advertise. However we look forward with some trepidation to the coming months and it is hard to imagine the world returning to anything that similar to what existed in January 2020. It is also hard to imagine GDC and MWC returning in 2021.
September 2020
