|Title||Advances in Cognitive Systems|
|Publication Type||Conference Proceedings|
|Year of Conference||2017|
|Authors||Roberts, M, Hiatt, LM|
|Conference Name||Improving Sequential Decision Making with Cognitive Priming|
Cognitive priming occurs when items or objects that are the focus of attention prime associated items in memory; these associations are learned over time between related items as an agent interacts with an environment. This mechanism guides attention toward associations relevant to the current situation and facilitates learning by retrieving those associated concepts from long-term memory. We apply priming to two sequential decision problems for the game of Minecraft: selecting the best tool to mine blocks and choosing the next subgoal in a maze. For learning tool selection, we leverage prior work that learned subgoal dependencies (e.g., wheat requires seeds) from the Minecraft community wiki. We apply these dependencies to train an associative memory that primes the selection of a tool for breaking a block and show that priming reduces the regret of choosing the incorrect tool. For subgoal selection, our prior work showed that decision trees can be learned to select the next subgoal for a simple variant of Minecraft. In two studies, we assess how well priming can select subgoals and show that priming performs comparably to decision trees. The first study shows that priming outperforms an incremental decision tree algorithm, while the second study examines the effect of noise in the training and testing data to show that priming performs comparably to a batch decision tree algorithm. Overall, these findings suggest that priming is an effective mechanism for either learning directly or for improving learning.
NRL Publication Release Number: