Plan Arithmetic: Compositional Plan Vectors for Multi-Task Control - ShortScience.org
If you've been at all aware of machine learning in the past five years, you've almost certainly seen the canonical word2vec example demonstrating additive properties of word embeddings: "king - man + woman = queen". This paper has a goal of designing embeddings for agent plans or trajectories that follow similar principles, such that a task composed of multiple subtasks can be represented by adding the vectors corresponding to the subtasks. For example, if a task involved getting an ax and then going to a tree, you'd want to be able to generate an embedding that corresponded to a policy to execute that task by summing the embeddings for "go to ax" and "go to tree". https://i.imgur.com/AHlCt76.png The authors don't assume that they know the discrete boundaries between subtasks in multiple-task trajectories, and instead use a relatively simple and clever training structure in order to induce the behavior described above. They construct some network g(x) that takes in information describing a trajectory (in this case, start and end state, but presumably could be more specific transitions), and produces an embedding. Then, they train a model on an imitation learning problem, where, given one demonstration of performing a particular task (typically generated by the authors to be composed of multiple subtasks), the agent needs to predict what action will be taken next in a second trajectory of the same composite task. At each point in the sequence of predicting the next action, the agent calculates the embedding of the full reference trajectory, and the embedding of the actions they have so far performed in the current stage in the predicted trajectory, and calculates the difference between these two values. This embedding difference is used to condition the policy function that predicts next action. At each point, you enforce this constraint, that the embedding of what is remaining to be done in the trajectory be close to the embedding of (full trajectory) - (what has so far been completed), by making the policy that corresponds with that embedding map to the remaining part of the trajectory. In addition to this core loss, they also have a few regularization losses, including: 1. A loss that goes through different temporal subdivisions of reference, and pushes the summed embedding of the two parts to be close to the embedding of the whole 2. A loss that simply pushes the embeddings of the two paired trajectories performing the same task closer together The authors test mostly on relatively simple tasks - picking up and moving sequences of objects with a robotic arm, moving around and picking up objects in a simplified Minecraft world - but do find that their central partial-conditioning-based loss gives them better performance on demonstration tasks that are made up of many subtasks. Overall, this is an interesting and clever paper: it definitely targeted additive composition much more directly, rather than situations like the original word2vec where additivity came as a side effect of other properties, but it's still an idea that I could imagine having interesting properties, and one I'd be interested to see applied to a wider variety of tasks.
Post to Tumblr