Catastrophic forgetting continues to severely restrict the learnability of controllers suitable for multiple task environments. Efforts to combat catastrophic forgetting reported in the literature to date have focused on how control systems can be updated more rapidly, hastening their adjustment from good initial settings to new environments, or more circumspectly, suppressing their ability to overfit to any one environment. When using robots, the environment includes the robot's own body, its shape and material properties, and how its actuators and sensors are distributed along its mechanical structure. Here we demonstrate for the first time how one such design decision (sensor placement) can alter the landscape of the loss function itself, either expanding or shrinking the weight manifolds containing suitable controllers for each individual task, thus increasing or decreasing their probability of overlap across tasks, and thus reducing or inducing the potential for catastrophic forgetting.

In domains where measures of utility for automatically-designed artefacts (or agents performing subjective tasks) are difficult or impossible to mathematically describe (such as ‘be interesting’), human interactive search algorithms are an attractive alternative. However, despite notable achievements, they are still designed around a specific search method, resulting in a lack of problem generality: applying a new search algorithm requires an excessive amount of redesign such that an altogether new interactive method is formed in the process. This leads to missed opportunities for human interactive methods to utilize the power of state of the art optimization algorithms. Here, we introduce for the first time a framework for human interactive optimization that is agnostic to both the search method and the application problem. Using 13 different search methods on 24 fitness functions commonly found in evolutionary algorithm benchmarks, we show that our approach works on the majority of tested applications: many of the search methods, provided with access to the fitness functions, performed no better than our framework, which employs surrogate human participants who act as less informed and erroneous representations of the fitness function. In this way, our framework for interactive optimization provides a scalable solution by facilitating the integration of numerous types of current state of the art or future search algorithms. Future work will involve generalizing this method to admit multi-objective optimization methods and validation with human participants.

Catastrophic interference occurs when an agent improves in one training instance but becomes worse in other instances. Many methods intended to combat interference have been reported in the literature that modulate properties of a neural controller, such as synaptic plasticity or modularity. Here, we demonstrate that adjustments to the body of the agent, or the way its performance is measured, can also reduce catastrophic interference without requiring changes to the controller. Additionally, we introduce new metrics to quantify catastrophic interference. We do not show that our approach outperforms others on benchmark tests. Instead, by more precisely measuring interactions between morphology, fitness, and interference, we demonstrate that embodiment is an important aspect of this problem. Furthermore, considerations into morphology and fitness can combine with, rather than compete with, existing methods for combating catastrophic interference.