Stephen Kelly

Stephen Kelly, Daniel S. Park, Xingyou Song et al.

Autonomous robots deployed in the real world will need control policies that rapidly adapt to environmental changes. To this end, we propose AutoRobotics-Zero (ARZ), a method based on AutoML-Zero that discovers zero-shot adaptable policies from scratch. In contrast to neural network adaptation policies, where only model parameters are optimized, ARZ can build control algorithms with the full expressive power of a linear register machine. We evolve modular policies that tune their model parameters and alter their inference algorithm on-the-fly to adapt to sudden environmental changes. We demonstrate our method on a realistic simulated quadruped robot, for which we evolve safe control policies that avoid falling when individual limbs suddenly break. This is a challenging task in which two popular neural network baselines fail. Finally, we conduct a detailed analysis of our method on a novel and challenging non-stationary control task dubbed Cataclysmic Cartpole. Results confirm our findings that ARZ is significantly more robust to sudden environmental changes and can build simple, interpretable control policies.

Stephen Kelly, Tatiana Voegerl, Wolfgang Banzhaf et al.

A fundamental aspect of behaviour is the ability to encode salient features of experience in memory and use these memories, in combination with current sensory information, to predict the best action for each situation such that long-term objectives are maximized. The world is highly dynamic, and behavioural agents must generalize across a variety of environments and objectives over time. This scenario can be modeled as a partially-observable multi-task reinforcement learning problem. We use genetic programming to evolve highly-generalized agents capable of operating in six unique environments from the control literature, including OpenAI's entire Classic Control suite. This requires the agent to support discrete and continuous actions simultaneously. No task-identification sensor inputs are provided, thus agents must identify tasks from the dynamics of state variables alone and define control policies for each task. We show that emergent hierarchical structure in the evolving programs leads to multi-task agents that succeed by performing a temporal decomposition and encoding of the problem environments in memory. The resulting agents are competitive with task-specific agents in all six environments. Furthermore, the hierarchical structure of programs allows for dynamic run-time complexity, which results in relatively efficient operation.