Biocomputers -- Applications to medicine, Chemical reactions -- Simulation methods, Reinforcement learning
Many potential applications of biochemical computers involve the detection of highly adaptable and dynamic chemical systems, such as emerging pathogens. Current technology is expensive to develop and unique to each application, thus causing limitations in accessibility. In order to make this type of computing a realistic solution to problems in the medical field, a biochemical computer would need to be adaptable to work in a variety of applications. Banda et al. (2014) previously proposed a first dynamic biochemical system that was capable of autonomous learning. For this project we studied a framework similar to Banda’s but in two separate pieces, an abstract chemical model and new methods of reinforcement learning.
An artificial chemistry can be used to simulate the behavior of a realistic chemistry. For the purpose of this project, a multiset of variables was used to represent concentrations of molecules. The behavior of the multiset was studied when applied with an Abstract Rewriting System on Multisets (ARMS). The ARMS is a set of rules that acts on the multiset for a number of terms. It was found that the concentrations of certain symbols over time create a cyclic pattern. Cycles of different patterns emerge under varying input rates, rule sets, and rule order.
A reinforcement learning algorithm called Q-learning was applied to a maze-like problem to study the learning’s effectiveness. A grid world was created with “food” at static locations to provide a reward. More results are needed to conclude the effectiveness of the learning algorithm that was used. Once this can be confirmed, an abstract chemistry will be used as the agent to be trained by reinforcement learning.
Braun, Lauren, "Learning in Bio-molecular Computing Systems" (2018). Undergraduate Research & Mentoring Program. 26.