Generating Heuristics for Novice Players

We consider the problem of generating compact sub-optimal game-playing heuristics that can be understood and easily executed by novices. In particular, we seek to find heuristics that can lead to good play while at the same time be expressed as fast and frugal trees or short decision lists. This has applications in automatically generating tutorials and instructions for playing games, but also in analyzing game design and measuring game depth. We use the classic game Blackjack as a testbed, and compare condition induction with the RIPPER algorithm, exhaustive-greedy search in statement space, genetic programming and axis-aligned search. We find that all of these methods can find compact well-playing heuristics under the given constraints, with axis-aligned search performing particularly well.

Publication:

Fernando de Mesentier Silva, Aaron Isaksen, Julian Togelius, Andy Nealen. “Generating Heuristics for Novice Players” Proceedings of Computational Intelligence and Games (CIG). IEEE (2016).

PDF link: http://julian.togelius.com/Silva2016Generating.pdf

Fig. 1. Fast and Frugal Trees are a type of binary decision tree used in bounded rationality theory. They are effective and easy to learn heuristic for humans. In this paper, we generate fast and frugal trees – also known as decision lists – for beginning strategy for Blackjack.

Fig. 1. Fast and Frugal Trees are a type of binary decision tree used in bounded rationality theory. They are effective and easy to learn heuristic for humans. In this paper, we generate fast and frugal trees – also known as decision lists – for beginning strategy for Blackjack.

 

 Fig. 2. A Fast and Frugal heuristic generated by one the methods, Axis-Aligned Search, used in the paper.

Fig. 2. A Fast and Frugal heuristic generated by one the methods, Axis-Aligned Search, used in the paper.

 

Fig. 3. Comparison of heuristic complexity vs fitness for the various heuristics generated in the paper.

Fig. 3. Comparison of heuristic complexity vs fitness for the various heuristics generated in the paper.

Events

No Upcoming Events