Back when I was running the development studio for Koios Works, I was in charge of building out the AI for our games. At first, we built some fairly straight forward approaches to planning, but the AI ended up being so scripted that replay of game levels just wasn’t fun. (Sound familiar?) For our second title, we did a major upgrade of our graphics engine, and in parallel, we decided to do a major upgrade of our AI.
We ended up selecting a genetic algorithm that essentially understands ‘good behavior’ vs. ‘bad behavior’ in selecting a ‘good’ solution to the problem being presented. For games, this is a fairly straight-forward problem – decide on the best move against the human adversary following the rules of the game. It did this based on an ‘evaluation function’ that is able to numerically rate a proposed solution.
The algorithm itself then simulates a genetic evolution to match up the very best solutions found to, hopefully, generate the next generation of even better solutions. (The details are somewhat too complex to describe here, but generally, you just need to focus on creating an evaluation that is able to rate options – all the hard work and intelligence is already coded into the algorithm itself.) In the end, our AI was considered among the very best developed for war-games (at least based on feedback from players in that fairly niche area).
To help bring in additional revenues during that time, I was also doing some outside consulting work. And, we were able to apply the same AI approach to a ship scheduling (i.e., supply chain) problem for the petroleum industry. The reason why you use this type of algorithm is that it is able to develop approximations of a great solution in a short amount of time while still being flexible to appreciate new variations to the problem (for example, the customer is this case was able to add in new ships or new pickup opportunities to perform what-if analysis – as well as insert new rules about profitability vs. customer satisfaction, etc.)
What got me thinking about this recently was some work we are doing for a client in the areas of digital supply chain. Just like the real world, when you are doing content processing from step-to-step, there are some real resource constraints that you should be aware of. From a human perspective, you have to balance the use of QC resources, touch-points for approvals, manual tasks (such as encoding a video or other master material into digital form), financial approvals, etc. On the computing side, you are potentially talking about huge amounts of disk and CPU resource utilization that can quickly translate into expensive infrastructure. To help optimize these resources, the application of AI is a great solution; and, a genetic algorithm specifically, is a nice fit to the problem since it can so easily be expanded with new influencers as well as adopt to variations in the type of content being processed (i.e., music vs. video vs. gaming, etc.)
So, while the emerging area of digital supply chains has different issues than traditional supply chain problems, let’s not forget to incorporate the lessons learned and optimization approaches that we’ve relied on for years.
By way of background, I originally wanted to become a neurologist when I went to college. My undergraduate degree was in Neuroethology mixed with a minor in Psychology. Go ahead, look up neuroethology, I’ll wait 
Actually, don’t look it up – I can sum it up as basically understanding behavior based on the way you are wired. So, John’s book smacks right into the type of stuff I really was interested in.