R12: Restricted Play

Read and respond to https://homes.cs.washington.edu/~zoran/jaffe2012ecg.pdf Links to an external site.

Towards the end, note this passage: "Our prototype uses optimal agents; it will be worthwhile to examine the usefulness of restricted play for strong but imperfect agents. This may be more realistic, and will be necessary for more complex games. We propose to exploit recent advances in heuristic AI - in particular, we are experimenting with Monte-CarloTree Search (MCTS)."

Behind-the-scenes note: The card game explored in this paper later evolved into and was deployed as Creature Capture (http://centerforgamescience.org/blog/portfolio/creature-capture/ Links to an external site.). The single-player version of Creature Capture initially involved playing against an opponent implemented with the vanilla UCT algorithm. The strength of this opponent was such that it outsmarted the (elementary grade-5 student) player to such a degree that one started crying in a playtest (or so a story in the lab goes). Reducing the number of samples used by UCT lead only to an inconsistently brutal opponent. We (I was in the lab at this time) later reformulated the opponent as an expectimax* search playing with the goal of being beaten by the human player only after they used a key move (modeling the human's strategy as essentially the random moves of nature).

* well, sorta. I think we averaged scores for the next moves over 10 random determinizations of the game.