In this paper, we explain the Partial Mental State Inducer (PMSI), which models how humans can
learn intuition for problem solving. We believe that this is the first model that unifies theories
of K-lines, human inner language, human problem solving, and neural reinforcement learning.
The result is human-like learning in a training environment that is too data-starved for traditional
reinforcement learning to be successful. We present experiments in three distinct problem domains
(natural language queries, equation solving, and robot planning) with only 20 training problems in
each domain. A typical deep Q-network set-up often does not test better than a random agent on
average, whereas PMSI can learn how to perform at a level that is close to optimal.
