The results of the experiments were not what were expected. The “learning” AI took longer and longer to “win” and the lessons that they learned were probably not desirable in the “real world.” See below:
Indeed, the experiment produced only six learned behaviors (see table above). The four learned behaviors for the “non-aggressive” tanks were to retreat. Unfortunately, the retreat was simply backing up so the tank was still hit by the incoming missile. The two learned behaviors for the aggressive tanks were if you see something on your radar, attack it.
Comments:
While KA is almost certainly an important tool in creating CGF AI, this experiment, in my opinion, is almost worthless.
First, one of the goals of the authors’ is, “To implement a truly autonomous agent, the agent on its own must encode the appropriate attributes which are relevant to the execution of behaviors. This remains a challenge for the simulation of human behavior and for furthering an understanding of learning.” However, human behavior, especially within a military context, is not simply “aggressive” or “nonaggressive.”
General S. L. A. Marshall did a great deal of research into this very subject which was published in his classic work, “Men Against Fire.” Marshall discovered that less than 25% of combatants in World War II could be classified as “self starters” (he believed the actual number was probably closer to 5%). The actions of these “self starters” determined the actions of the rest of the troops in their immediate vicinity. If they moved forward, the rest of the platoon moved forward. If they were killed or were incapacitated, the rest of the platoon ground to a halt. Marshall’s work greatly influenced my equations used in my wargames to determine when and why a unit retreated.
Second, the game TankSoar is far too simple to have any military value whatsoever. Indeed, while the experimenters are using terms like “tank”, “missile” and “radar” they could have just as well have been “pawns” or “pieces”. The tanks did not have any of the characteristics of tanks and the “battlefield” bore no resemblance to real terrain whatsoever.
Lastly, while the “tanks” did, indeed, learn some lessons, they were either not the lessons that you would have wanted them to learn in real life (in the words of Monty Python, “run away!”) or so absurdly obvious (go towards enemy units) that any CGF AI should have been programmed to do this in the first place.
While the authors claim the experiment was a success (“The proposed approach was able to allow Soar CGFs to learn autonomously over successive trials and to The Current State of Human-Level Artificial Intelligence in Computer Simulations and Wargames. Page 66 of 116 adaptively add and drop rules according to their experience with little human involvement. These adaptations were achieved efficiently at a very fast learning rate.”) any real CGF AI using these methods are considerably further down this research road.
