Instrumental Conditioning with Neuromodulated Plasticity on SpiNNaker.

We present a work-in-progress on implementing reinforcement learning by instrumental conditioning on SpiNNaker. Animals learn to behave by exploring the changing environment around them such that, over a period of time, their behaviour gives a good outcome (reward) i.e. a perception of ‘satisfaction’. While inspired by animal learning, reinforcement learning adopts a goal-directed strategy of maximising rewards in a dynamic environment. Instrumental conditioning is a strategy to strengthen the association between an action and the environmental state when the state-action pair is rewarded i.e. the reward is instrumental in forming the association. However, in the real world, the delivery of a reward is often delayed in time, known as the distal reward problem. Using the concept of eligibility traces and spike-time dependant plasticity (STDP), Izhikevich (2007) simulated both classical and instrumental conditioning in a spiking neural network with Dopamine (DA)-modulated STDP. The current implementation of DA-modulated plasticity on SpiNNaker using trace-based STDP is reported by Mikaitas et al. (2018), who demonstrated classical conditioning with a similar experimental set up as Izhikevich. Our results show that using delayed DA-modulation of STDP on SpiNNaker, we can condition a neural population to maximise its reward over a period of time by firing at a higher rate than another competing population. Ongoing work is looking into a dynamic conditioning scenario where different actions can be selected within the same run as is the case in real world scenarios.