The Evolution of Learning: Balancing adaptivity and stability in artificial agents
MetadataVis full innførsel
A longstanding challenge in artificial intelligence is to create agents that learn, enabling them to interact with and adapt to a complex and changing world. A better understanding of the evolution of learning may help produce robust and adaptive agents, as well as shed light on open questions about the evolution of learning from biology. Evolutionary computation offers the benefits of precise experimental control, repeatability of experiments and rapid generational turnover – enabling experiments to test hypotheses that would be impossible or extremely time demanding to test in natural studies. The evolution of learning is influenced by the balance between the benefits offered by adaptivity and the costs (disadvantages) individuals pay for learning abilities. Such costs include forgetting previous knowledge, dangers of exploration and maintenance of neural structures for learning. This thesis focuses on how evolution regulates learning capacities to reap the benefits of being adaptive, while minimizing the costs of learning. The regulation of learning capacities is studied along three main axes: regulation through individual lifetimes, regulation within a population facing varying environments and regulation across neural modules. The study of learning regulation within individual lifetimes is inspired by the sensitive periods in learning observed in nature: limited periods within individuals’ lives where learning is temporarily facilitated. Experiments herein demonstrate that sensitive periods can emerge to schedule learning in tasks where there are dependencies between the learning of sub-tasks, and further explore how the flexibility of evolved sensitive periods depends on assumptions about which factors regulate plasticity. On the population level, the evolution of learning efforts is known to be highly dependent on the variability of the environment and the reliability of environmental stimuli. Evolving the innate preferences and learning rates of individuals across a wide range of environmental variability demonstrates that environments changing too rapidly or too slowly discourage the evolution of learning. Further experiments show how independently varying the degrees of environmental stability and stimuli reliability leads to a refinement of this model of learning, which also acknowledges the fact that learning may be disruptive or inefficient when stimuli are not reliable. One cost of learning is the risk of losing old information as new information is gained, a problem known as catastrophic forgetting. Evolving individuals facing a task with potential for catastrophic forgetting, it is demonstrated how the addition of an evolutionary cost of neural connections leads to more modular networks, which forget old skills less when learning a new skill. Together, the findings herein demonstrate several ways to handle the so-called stabilityplasticity dilemma: how can an individual be realized which has the flexibility to adapt without risking unstable behaviors and forgetting of old skills? The findings suggest ways in which evolution may have solved this problem in natural learners, and ways to harness the powers of evolution to mitigate this problem in artificial agents.