Summary of Planning in a Recurrent Neural Network That Plays Sokoban, by Mohammad Taufeeque et al.
Planning in a recurrent neural network that plays Sokoban
by Mohammad Taufeeque, Philip Quirke, Maximilian Li, Chris Cundy, Aaron David Tucker, Adam Gleave, Adrià Garriga-Alonso
First submitted to arxiv on: 22 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary An investigation into neural networks’ ability to generalize to novel situations reveals that recurrent neural networks (RNNs) trained for tasks like Sokoban appear to use a planning process. This is evident from the extra computation steps required to solve complex scenarios, which allows the RNN to “pace” itself in cycles. By analyzing the hidden state of the network, we can predict future actions taken by the agent and control its subsequent decisions using linear probes. Building on these insights, we modify a convolutional neural network (CNN) to generalize beyond its architectural limit, enabling it to solve challenging levels with arbitrarily sized inputs. Our model, which has only 1.29 million parameters, is an ideal testbed for studying learned planning. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine you have a special kind of computer program called a neural network that can learn to do tasks like solving puzzles or playing games. Researchers wanted to know how these programs can generalize, or apply what they’ve learned, to new and more challenging situations. They found that some neural networks use a planning process to make decisions, which helps them solve complex problems. By understanding how this works, the researchers were able to improve the performance of another type of neural network, called a convolutional neural network. This improved model can now solve very difficult levels with inputs of any size. The team is sharing their work and hopes it will help us better understand how neural networks learn and make decisions. |
Keywords
» Artificial intelligence » Cnn » Neural network » Rnn
