Learning near-optimal behaviour from an expert's demonstrations typically relies on the assumption that the learner knows the features that the true reward function depends on. In this paper, we study the problem of learning from demonstrations in the setting where this is not the case, i.e., where there is a mismatch between the worldviews of the learner and the expert. We introduce a natural quantity, the teaching risk, which measures the potential suboptimality of policies that look optimal to the learner in this setting. We show that bounds on the teaching risk guarantee that the learner is able to find a near-optimal policy using standard algorithms based on inverse reinforcement learning. Based on these findings, we suggest a teaching scheme in which the expert can decrease the teaching risk by updating the learner's worldview, and thus ultimately enable her to find a near-optimal policy.

Crowd simulators have been used to find appropriate regulations by simulating multiple scenarios with different regulations. However, this approach requires multiple simulation runs, which are time-consuming. In this paper, we propose a method to learn a function that outputs regulation effects given the current traffic situation as inputs. If the function is learned using the training data of many simulation runs in advance, we can obtain an appropriate regulation efficiently by bypassing simulations for the current situation. We use the graph convolutional networks for modeling the function, which enable us to find regulations even for unseen areas. With the proposed method, we construct a graph for each area, where a node represents a road, and an edge represents the road connection. By running crowd simulations with various regulations on various areas, we generate traffic situations and regulation effects. The graph convolutaional networks are trained to output the regulation effects given the graph with the traffic situation information as inputs. With experiments using real-world road networks and a crowd simulator, we demonstrate that the proposed method can find a road to close that reduces the average time needed to reach the destination.

We explore Deep Reinforcement Learning in a parameterized action space. Specifically, we investigate how to achieve sample-efficient end-to-end training in these tasks. We propose a new compact architecture for the tasks where the parameter policy is conditioned on the output of the discrete action policy. We also propose two new methods based on the state-of-the-art algorithms Trust Region Policy Optimization (TRPO) and Stochastic Value Gradient (SVG) to train such an architecture. We demonstrate that these methods outperform the state of the art method, Parameterized Action DDPG, on test domains.

We present a framework, which we call Molecule Deep $Q$-Networks (MolDQN), for molecule optimization by combining domain knowledge of chemistry and state-of-the-art reinforcement learning techniques (prioritized experience replay, double $Q$-learning, and randomized value functions). We directly define modifications on molecules, thereby ensuring 100% chemical validity. Further, we operate without pre-training on any dataset to avoid possible bias from the choice of that set. As a result, our model outperforms several other state-of-the-art algorithms by having a higher success rate of acquiring molecules with better properties. Inspired by problems faced during medicinal chemistry lead optimization, we extend our model with multi-objective reinforcement learning, which maximizes drug-likeness while maintaining similarity to the original molecule. We further show the path through chemical space to achieve optimization for a molecule to understand how the model works.

Today's topic is … well, the same as the last one. Last time, we explained what Q Learning is and how to use the Bellman equation to find the Q-values and as a result the optimal policy. Later, we introduced Deep Q Networks and how instead of computing all the values of the Q-table, we let a Deep Neural Network learn to approximate them. Deep Q Networks take as input the state of the environment and output a Q value for each possible action. The maximum Q value determines, which action the agent will perform. The training of the agents uses as loss the TD Error, which is the difference between the maximum possible value for the next state and the current prediction of the Q-value (as the Bellman equation suggests).

We are excited to announce the release of the Deep Learning and Reinforcement Learning Summer School videos!

The classic objective in a reinforcement learning (RL) problem is to find a policy that minimizes, in expectation, a long-run objective such as the infinite-horizon discounted or long-run average cost. In many practical applications, optimizing the expected value alone is not sufficient, and it may be necessary to include a risk measure in the optimization process, either as the objective or as a constraint. Various risk measures have been proposed in the literature, e.g., mean-variance tradeoff, exponential utility, the percentile performance, value at risk, conditional value at risk, prospect theory and its later enhancement, cumulative prospect theory. In this article, we focus on the combination of risk criteria and reinforcement learning in a constrained optimization framework, i.e., a setting where the goal to find a policy that optimizes the usual objective of infinite-horizon discounted/average cost, while ensuring that an explicit risk constraint is satisfied. We introduce the risk-constrained RL framework, cover popular risk measures based on variance, conditional value-at-risk and cumulative prospect theory, and present a template for a risk-sensitive RL algorithm. We survey some of our recent work on this topic, covering problems encompassing discounted cost, average cost, and stochastic shortest path settings, together with the aforementioned risk measures in a constrained framework. This non-exhaustive survey is aimed at giving a flavor of the challenges involved in solving a risk-sensitive RL problem, and outlining some potential future research directions.

We propose an efficient multi-agent reinforcement learning approach to derive equilibrium strategies for multi-agents who are participating in a Markov game. Mainly, we are focused on obtaining decentralized policies for agents to maximize the performance of a collaborative task by all the agents, which is similar to solving a decentralized Markov decision process. We propose to use two different policy networks: (1) decentralized greedy policy network used to generate greedy action during training and execution period and (2) generative cooperative policy network (GCPN) used to generate action samples to make other agents improve their objectives during training period. We show that the samples generated by GCPN enable other agents to explore the policy space more effectively and favorably to reach a better policy in terms of achieving the collaborative tasks.

Value-based approaches can be difficult to use in continuous action spaces, because an optimization has to be solved to find the greedy action for the action-values. A common strategy has been to restrict the functional form of the action-values to be convex or quadratic in the actions, to simplify this optimization. Such restrictions, however, can prevent learning accurate action-values. In this work, we propose the Actor-Expert framework for value-based methods, that decouples action-selection (Actor) from the action-value representation (Expert). The Expert uses Q-learning to update the action-values towards the optimal action-values, whereas the Actor (learns to) output the greedy action for the current action-values. We develop a Conditional Cross Entropy Method for the Actor, to learn the greedy action for a generically parameterized Expert, and provide a two-timescale analysis to validate asymptotic behavior. We demonstrate in a toy domain with bimodal action-values that previous restrictive action-value methods fail whereas the decoupled Actor-Expert with a more general action-value parameterization succeeds. Finally, we demonstrate that Actor-Expert performs as well as or better than these other methods on several benchmark continuous-action domains.

We propose a practical non-episodic PSRL algorithm that unlike recent state-of-the-art PSRL algorithms uses a deterministic, model-independent episode switching schedule. Our algorithm termed deterministic schedule PSRL (DS-PSRL) is efficient in terms of time, sample, and space complexity. We prove a Bayesian regret bound under mild assumptions. Our result is more generally applicable to multiple parameters and continuous state action problems. We compare our algorithm with state-of-the-art PSRL algorithms on standard discrete and continuous problems from the literature. Finally, we show how the assumptions of our algorithm satisfy a sensible parametrization for a large class of problems in sequential recommendations.