A longstanding challenge is to make intelligent agents reason more abstractly and farther ahead; in other words, to think strategically. The difficulty is that traditional approaches to planning focus on the effects of discrete, ground actions. High-level abstract goals may not be amenable to projecting concrete future states, yet they provide important long-term guidance in selecting actions, allocating resources, and focusing attention. We are exploring the representation and exploitation of strategy as a qualitative modeling problem. Our key hypothesis is that the same reasoning mechanisms can be applied to quantities at all levels of abstraction and that strategies can be expressed with respect to a domain-neutral vocabulary of processes, states, and goal tradeoffs in a fashion that is learnable, expressible, and explainable. We illustrate these ideas in the domain of Freeciv, an open-source strategy game.
|Original language||English (US)|
|Number of pages||18|
|Journal||Advances in Cognitive Systems|
|State||Published - 2016|