Low-pass Recurrent Neural Networks - A memory architecture for longer-term correlation discovery
This addresses memory and learning efficiency issues in reinforcement learning for agents performing complex tasks, but it appears incremental as it builds on existing backpropagation through time methods.
The paper tackles the problem of reinforcement learning agents needing to remember observations and actions over long time intervals, especially during exploration, by proposing a low-pass recurrent neural network memory architecture that extends the effective window for backpropagation through time without requiring longer traces, showing empirical results on a few tasks.
Reinforcement learning (RL) agents performing complex tasks must be able to remember observations and actions across sizable time intervals. This is especially true during the initial learning stages, when exploratory behaviour can increase the delay between specific actions and their effects. Many new or popular approaches for learning these distant correlations employ backpropagation through time (BPTT), but this technique requires storing observation traces long enough to span the interval between cause and effect. Besides memory demands, learning dynamics like vanishing gradients and slow convergence due to infrequent weight updates can reduce BPTT's practicality; meanwhile, although online recurrent network learning is a developing topic, most approaches are not efficient enough to use as replacements. We propose a simple, effective memory strategy that can extend the window over which BPTT can learn without requiring longer traces. We explore this approach empirically on a few tasks and discuss its implications.