Mental representations, or sensor-independent internal models of the environment, are used to interpret the world and make decisions based upon that understanding. For example, a human sees dark clouds in the sky, recalls that often dark clouds mean rain (a mental representation), and consequently decides to wear a raincoat. I seek to identify, understand, and encourage the evolution of these representations in silico. Previous work identified an information-theoretic tool, referred to as R, that measures mental representations in artificial cognitive systems (e.g., Markov Brains or Recurrent Neural Networks). Further work found that selecting for R, along with task performance, in the evolution of artificial cognitive systems leads to better overall performance on a given task. Here I explore the implications and opportunities of this modified selection process, referred to as R-augmentation. After an overview of common methods, techniques, and computational substrates in Chapter 1, a series of working chapters experimentally demonstrate the capabilities and possibilities of R-augmentation. First, in Chapter 2, I address concerns regarding potential limitations of R-augmentation. This includes an refutation of suspected negative impacts on the system’s ability to generalize within-domain and the system’s robustness to sensor noise. To the contrary, the systems evolved with R-augmentation tend to perform better than those evolved without, in the context of noisy environments and different computational components. In Chapter 3 I examine how R-augmentation works across different cognitive structures, focusing on the evolution of genetic programming related structures and the effect that augmentation has on the distribution of their representations. For Chapter 4, in the context of the all-component Markov Brain (referred to as a Buffet Brain, see [Hintze et al., 2019]) I analyze potential reasons that explain why R-augmentation works; the mechanism seems to be based on evolutionary dynamics as opposed to structural or component differences. Next, I demonstrate a novel usage of R-augmentation in Chapter 5; with R-augmentation, one can use far fewer training examples during evolution and the resulting systems still perform approximately as well as those that were trained on the full set of examples. This advantage in increased performance at low sample size is found in some examples of in-domain and out-domain generalization, with the ”worst-case” scenario being that the networks created by R-augmentation perform as well as their unaugmented equivalents. Lastly, in Chapter 6 I move beyond R-augmentation to explore using other neuro-correlates - particularly the distribution of representations, called smearedness - as part of the fitness function. I investigate the possibility of using MAP-Elites to identify an optimal value of smearedness for augmentation or for use as an optimization method in its own right. Taken together, these investigations demonstrate both the capabilities and limitations of R-augmentation, and open up pathways for future research.
Read
