Explaining Decisions of a Deep Reinforcement Learner with a Cognitive Architecture
The work presented is an evaluation of a method for developing a hybrid system, consisting of a Deep Reinforcement Learning (RL) agent and a cognitive model, capable of providing explanations of its action decisions. The methodology uses a symbolic/sub-symbolic cognitive architecture to introspect on the activity of the network to understand its representation. The entropy in the system’s behavioral predictions could be used as a signal to affirm or deny ascribing a representation to the network.
deep reinforcement learning, cognitive modeling, introspection
Somers, Sterling; Mitsupoulos, Constantinos; Lebiere, Christian; and Thomson, Robert, "Explaining Decisions of a Deep Reinforcement Learner with a Cognitive Architecture" (2018). ACI Journal Articles. 124.