Timezone: »

Poster
No-Regret Exploration in Goal-Oriented Reinforcement Learning
Jean Tarbouriech · Evrard Garcelon · Michal Valko · Matteo Pirotta · Alessandro Lazaric

Wed Jul 15 12:00 PM -- 12:45 PM &amp; Thu Jul 16 01:00 AM -- 01:45 AM (PDT) @
Many popular reinforcement learning problems (e.g., navigation in a maze, some Atari games, mountain car) are instances of the episodic setting under its stochastic shortest path (SSP) formulation, where an agent has to achieve a goal state while minimizing the cumulative cost. Despite the popularity of this setting, the exploration-exploitation dilemma has been sparsely studied in general SSP problems, with most of the theoretical literature focusing on different problems (i.e., fixed-horizon and infinite-horizon) or making the restrictive loop-free SSP assumption (i.e., no state can be visited twice during an episode). In this paper, we study the general SSP problem with no assumption on its dynamics (some policies may actually never reach the goal). We introduce UC-SSP, the first no-regret algorithm in this setting, and prove a regret bound scaling as $\widetilde{\mathcal{O}}( D S \sqrt{ A D K})$ after $K$ episodes for any unknown SSP with $S$ states, $A$ actions, positive costs and SSP-diameter $D$, defined as the smallest expected hitting time from any starting state to the goal. We achieve this result by crafting a novel stopping rule, such that UC-SSP may interrupt the current policy if it is taking too long to achieve the goal and switch to alternative policies that are designed to rapidly terminate the episode.

#### Author Information

##### Michal Valko (DeepMind)

Michal is a research scientist in DeepMind Paris and SequeL team at Inria Lille - Nord Europe, France, lead by Philippe Preux and Rémi Munos. He also teaches the course Graphs in Machine Learning at l'ENS Cachan. Michal is primarily interested in designing algorithms that would require as little human supervision as possible. This means 1) reducing the “intelligence” that humans need to input into the system and 2) minimising the data that humans need spend inspecting, classifying, or “tuning” the algorithms. Another important feature of machine learning algorithms should be the ability to adapt to changing environments. That is why he is working in domains that are able to deal with minimal feedback, such as bandit algorithms, semi-supervised learning, and anomaly detection. Most recently he has worked on sequential algorithms with structured decisions where exploiting the structure can lead to provably faster learning. In the past the common thread of Michal's work has been adaptive graph-based learning and its application to the real world applications such as recommender systems, medical error detection, and face recognition. His industrial collaborators include Adobe, Intel, Technicolor, and Microsoft Research. He received his PhD in 2011 from University of Pittsburgh under the supervision of Miloš Hauskrecht and after was a postdoc of Rémi Munos.