Cognitive policy learner: biasing winning or losing strategies

In continuous learning settings stochastic stable policies are often necessary to ensure that agents continuously adapt to dynamic environments. The choice of the decentralised learning system and the employed policy plays an important role in the optimisation task. For example, a policy that exhibits fluctuations may also introduce non-linear effects which other agents in the environment may not be able to cope with and even amplify these effects. In dynamic and unpredictable multiagent environments these oscillations may introduce instabilities. In this paper, we take inspiration from the limbic system to introduce an extension to the weighted policy learner, where agents evaluate rewards as either positive or negative feedback, depending on how they deviate from average expected rewards. Agents have positive and negative biases, where a bias either magnifies or depresses a positive or negative feedback signal. To contain the non-linear effects of biased rewards, we incorporate a decaying memory of past positive and negative feedback signals to provide a smoother gradient update on the probability simplex, spreading out the effect of the feedback signal over time. By splitting the feedback signal, more leverage on the win or learn fast (WoLF) principle is possible. The cognitive policy learner is evaluated using a small queueing network and compared with the fair action and weighted policy learner. Emphasis is placed on analysing the dynamics of the learning algorithms with respect to the stability of the queueing network and the overall queueing performance.