Traditional reinforcement learning (RL) populations evolve risk-aversion. Recently, the concept of nurturing has added another layer of complexity and possibility to artificial neural networks. Ideally, an RL population would evolve risk-aversion on its own, and when augmented with a safe-exploration nurturing period, would instead evolve risk-neutrality. My experiment specifically shows that in an environment with two risky options and one safe option, an asymmetric reward distribution causes nurtured agents to learn to take risks and distinguish between the different risky options, taking over the niche and pushing the non-nurtured population out or to extinction. This has applications in classifying environments based on whether they will support risky or non-risky populations, based solely on the setup parameters. This model can be used to evolve risk evaluation software for AI or to predict future trends in a biological population based on the current "riskiness" of its agents.
When I first started this project, I learned that computer science is not as finite a field as I had before assumed. When designing simulations, there is inevitably variability and randomness, but this randomness can be grouped into trends, forming order out of chaos. I also learned that although words like “machine learning” and “neural network” sound like advanced technical jargon, getting into computer science research is not nearly as hard as I expected. I was in my first semester at OU when I applied for the HERE program in Dr. Hougen’s REAL Lab, and after working in his lab for two semesters now I can say that I understand all the theory well enough to make it work with code, and I’m on the path to truly grasping the theory as well. Working on my projects in short bursts while pondering my next steps in between has proven a great strategy for not wasting my energy on future problems I may have already unwittingly solved.
Presented at Summer 2018 Curiosity 2 Creativity Symposium