Hi there. I'd been hoping in the last few weeks to deliver an update to Undaunted Normandy, but this isn't quite possible right now so I just wanted to give a status update.
The last release contained a significant improvement on the AI implementation, built using an AI self training solution. The intelligence of this solution is far higher than something that I could write when completed with a full training programme. Following the successful implementation of this for scenario 1 I started training it for scenario 2.
Training for scenario 1 shows that the German side wins approximately 2/3rds of all games (so if you want a greater challenge play as the US). However, in training for scenario 2 (where the US is trying to capture an objective and the Germans are trying to pin the US) the training process resulted in an almost 0% success rate for the German side. This was due to the training solution learning in a scenario with very differing objectives and not being able to cope. This also caused the training libraries to crash a lot.
The solution to this has been to implement two separate training agents in the environment, one to learn how to play as the US team and one to learn how to play as the German team. As a result, the individual agents are able to learn the appropriate strategies for their own approach to each scenario.
As part of the testing for this approach I've gone back to scenario 1 to run this processes, where the training is taking place (since even for this scenario the approach to winning is unique for each side). The result being that for all scenarios the AI opponent will be tuned to whichever team you're playing against.
The result of this will be a much better AI. My own long held belief (which I've written about here before) is that for ANY strategy game, the quality of AI is the number one, number two and number three priority and I'm completely committed to this. The downside is that training two sides to play a scenario individually takes a bit longer than getting one solution to play both sides, but good progress is being made. But on the upside the training libraries are much more stable in this way and almost never crash (so I can leave it running overnight).
So thats where we are right now - training for scenario 1 is still running but it's looking good and I hope to have a release out in a couple of weeks with a couple of other minor fixes, and scenario 2 training up and running again. Plus a solution which will work for training future scenarios.
Thanks for reading!