Inductive Biases, Invariances and Generalization in Reinforcement Learning

Anirudh Goyal, Rosemary Nan Ke, Jane Wang, Theo Weber, Fabio Viola, Bernhard Schölkopf, Stefan Bauer

Keywords:  reinforcement learning    inductive bias    generalization  


One proposed solution towards the goal of designing machines that can extrapolate experience across environments and tasks, are inductive biases. Providing and starting algorithms with inductive biases might help to learn invariances e.g. a causal graph structure, which in turn will allow the agent to generalize across environments and tasks.
While some inductive biases are already available and correspond to common knowledge, one key requirement to learn inductive biases from data seems to be the possibility to perform and learn from interventions. This assumption is partially motivated by the accepted hypothesis in psychology about the need to experiment in order to discover causal relationships. This corresponds to an reinforcement learning environment, where the agent can discover causal factors through interventions and observing their effects.

We believe that one reason which has hampered progress on building intelligent agents is the limited availability of good inductive biases. Learning inductive biases from data is difficult since this corresponds to an interactive learning setting, which compared to classical regression or classification frameworks is far less understood e.g. even formal definitions of generalization in RL have not been developed. While Reinforcement Learning has already achieved impressive results, the sample complexity required to achieve consistently good performance is often prohibitively high. This has limited most RL to either games or settings where an accurate simulator is available. Another issue is that RL agents are often brittle in the face of even tiny changes to the environment (either visual or mechanistic changes) unseen in the training phase.

To build intuition for the scope of the generalization problem in RL, consider the task of training a robotic car mechanic that can diagnose and repair any problem with a car. Current methods are all insufficient in some respect -- on-policy policy gradient algorithms need to cycle through all possible broken cars on every single iteration, off-policy algorithms end up with a mess of instability due to perception and highly diverse data, and model-based methods may struggle to fully estimate a complex web of causality.

In our workshop we hope to explore research and new ideas on topics related to inductive biases, invariances and generalization, including:

- What are efficient ways to learn inductive biases from data?
- Which inductive biases are most suitable to achieve generalization?
- Can we make the problem of generalization in particular for RL more concrete and figure out standard terms for discussing the problem?
- Causality and generalization especially in RL
- Model-based RL and generalization.
- Sample Complexity in reinforcement learning.
- Can we create models that are robust visual environments, assuming all the underlying mechanics are the same. Should this count as generalization or transfer learning?
- Robustness to changes in the mechanics of the environment, such as scaling of rewards.
- Can we create a theoretical understanding of generalization in RL, and understand how it is related to the well developed ideas from statistical learning theory.
- in RL, the training data is collected by the agent and it is affected by the agent's policy.
Therefore, the training distribution is not a fixed distribution. How does this affect how we should think about generalization?

The question of generalization in reinforcement learning is essential to the field’s future both in theory and in practice. However there are still open questions about the right way to think about generalization in RL, the right way to formalize the problem, and the most important tasks. This workshop would help to address this issue by bringing together researchers from different backgrounds to discuss these challenges.

Live content is unavailable. Are you logged in?

Timezone: »


Are you logged in?