Transfer / meta / lifelong learning
-
RL with policy advice. Azar et al., ECML 2013.
- Reduction from RL to bandit problem. -
Regret bounds: sum of differences between actual policy and optimal policy.
-
Regret scales with the number of tasks \sqrt(M), rather than the state and action space.
-
Brunskill and Li, UAI 2013. Reduce from RL to (active) classification problem.
-
Provably speeding multitask RL. Guo and Brunskill, AAAI 2015. K tasks sampled from M tasks. Evaluation goal: provably improve performance. Approach: quickly cluster, then share.
-
Killian et al., NIPS 2017. Bayesian NNs for modeling MDP dynamics.
-
Smooth latent policy space for crossdomain transfer. Anmar et al., IJCAI 2015. Limited theoretical results (some nice convergence results).
-
Model-agnostic meta-learning. Finn et al., ICML 2017.