As a mathematical engineer, Kenji Doya approaches the goal of describing the most intricate brain mechanisms from a computational perspective. He constructs models of reinforcement learning involving the networked structures of the basal ganglia. His efforts are captured and expressed quantitatively as probabilities, regressions, and algorithms.

In this presentation, Doya covers basic concepts of reinforcement learning, then surveys the last decade of inquiry into the components of the basal ganglia circuit governing voluntary motion. Among the topics: action values, action candidates, and reward prediction involving the neurotransmitter dopamine; model-free versus model-based learning strategies; and the essential role of serotonin as modulator in the complex information loop.

Doya’s recent research is carried out via robots he calls “cyber rodents.” His dream as an undergraduate was to “build a robot that learns the variety of behaviors on its own.” That is, the computer, not the human engineer, teaches the robot to move. He accomplished this in designing a machine-creature exhibiting emotion-like attributes characterized as “depression,” “impulsivity,” “greed,” and “patience.”

Doya believes the “metaparameters” of reinforcement learning must be “tuned appropriately…Otherwise the performance of your learning is very, very poor.” The iterative process involves three terms -- the reward itself, the expected reward for a new state based on choice of action, and memory of the reward gained in the previous state. In the comparison, any differential greater than zero can be exploited for learning. The tradeoff: “No pain, no gain.”

As research advanced to increasing levels of structural specificity, Doya posited that “there seems to be spatial segregation in the function” of basal ganglia components. Specialization in aspects of reinforcement learning is now seen, for instance, in ventral versus dorsal areas of the striatum.

Differentiation is also found in the cortico-basal ganglia information network: not a simple closed loop, but parallel electrical pathways conducting distinct neural operations. Further, the neuromodulators each have their respective missions. Dopamine encodes the temporal difference error -- the reward learning signal. Acetylcholine affects learning rate through memory updates of actions and rewards. Noradrenaline controls width or randomness of exploration. Serotonin is implicated in “temporal discounting,” evaluating if a given action is worth the expected reward. Doya reminds us that clinically “it is well known that the serotonin function is impaired in the depression patient.”

The system of basal ganglia components and neuromodulators requires dynamic balancing. A delicate interplay determines outcomes for learning, actions, and affective states. Doya’s synthetic models are proxies for human behavior, and his computational framework describing the moving parts ultimately has therapeutic implications for psychiatric and neurological disorders.