Learning to rank (LTR) is an important artificial intelligence (AI) approach supporting the operation of many search engines. In large-scale search systems, the ranking results are continually improved with the introduction of more factors to be considered by LTR. However, the more factors being considered, the more computation resources required, which in turn, results in increased system response latency. Therefore, removing redundant factors can significantly improve search engine efficiency. In this paper, we report on our experience incorporating our Contextual Factor Selection (CFS) deep reinforcement learning approach into the Taobao e-commerce platform to optimize the selection of factors based on the context of each search query to simultaneously maintaining search result quality while significantly reducing latency. Online deployment on Taobao.com demonstrated that CFS is able to reduce average search latency under everyday use scenarios by more than 40% compared to the previous approach with comparable search result quality. Under peak usage during the Single’s Day Shopping Festival (November 11th) in 2017, CFS reduced the average search latency by 20% compared to the previous approach.

In large-scale search systems, the quality of the ranking results is continually improved with the introduction of more factors from complex procedures. Meanwhile, the increase in factors demands more computation resources and increases system response latency. It has been observed that, under some certain context a search instance may require only a small set of useful factors instead of all factors in order to return high quality results. Therefore, removing ineffective factors accordingly can significantly improve system efficiency. In this paper, we report our experience incorporating our Contextual Factor Selection (CFS) approach into the Taobao e-commerce platform to optimize the selection of factors based on the context of each search query in order to simultaneously achieve high quality search results while significantly reducing latency time. This problem is treated as a combinatorial optimization problem which can be tackled through a sequential decision-making procedure. The problem can be efficiently solved by CFS through a deep reinforcement learning method with reward shaping to address the problems of reward signal scarcity and wide reward signal distribution in real-world search engines. Through extensive offline experiments based on data from the Taobao.com platform, CFS is shown to significantly outperform state-of-the-art approaches. Online deployment on Taobao.com demonstrated that CFS is able to reduce average search latency time by more than 40% compared to the previous approach with negligible reduction in search result quality. Under peak usage during the Single's Day Shopping Festival (November 11th) in 2017, CFS reduced peak load search latency time by 33% compared to the previous approach, helping Taobao.com achieve 40% higher revenue than the same period during 2016.

In industrial large-scale search systems, such as Taobao.com search for commodities, the quality of the ranking result is getting continually improved by introducing more factors from complex procedures, e.g., deep neural networks for extracting image factors. Meanwhile, the increasing of the factors demands more computation resource and raises the system response latency. It has been observed that a search instance usually requires only a small set of effective factors, instead of all factors. Therefore, removing ineffective factors significantly improves the system efficiency. This paper studies the \emph{Contextual Factor Selection} (CFS), which selects only a subset of effective factors for every search instance, for a well balance between the search quality and the response latency. We inject CFS into the search engine ranking score to accelerate the engine, considering both ranking effectiveness and efficiency. The learning of the CFS model involves a combinatorial optimization, which is transformed as a sequential decision-making problem. Solving the problem by reinforcement learning, we propose the RankCFS, which has been assessed in an off-line environment as well as a real-world on-line environment (Taobao.com). The empirical results show that, the proposed CFS approach outperforms several existing supervised/unsupervised methods for feature selection in the off-line environment, and also achieves significant real-world performance improvement, in term of service latency, in daily test as well as Singles' Day Shopping Festival in $2017$.

The success or failure of any learning algorithm is partially due to the exploration strategy it exerts. However, most exploration strategies assume that the environment is stationary and non-strategic. In this work we shed light on how to design exploration strategies in non-stationary and adversarial environments. Our proposed adversarial drift exploration (DE) is able to efficiently explore the state space while keeping track of regions of the environment that have changed. This proposed exploration is general enough to be applied in single agent non-stationary environments as well as in multiagent settings where the opponent changes its strategy in time. We use a two agent strategic interaction setting to test this new type of exploration, where the opponent switches between different behavioral patterns to emulate a non-deterministic, stochastic and adversarial environment. The agent’s objective is to learn a model of the opponent’s strategy to act optimally. Our contribution is twofold. First, we present DE as a strategy for switch detection. Second, we propose a new algorithm called R-max# for learning and planning against non-stationary opponent. To handle such opponents, R-max# reasons and acts in terms of two objectives: (1) to maximize utilities in the short term while learning and (2) eventually explore opponent behavioral changes. We provide theoretical results showing that R-max# is guaranteed to detect the opponent’s switch and learn a new model in terms of finite sample complexity. R-max# makes efficient use of exploration experiences, which results in rapid adaptation and efficient DE, to deal with the non-stationary nature of the opponent. We show experimentally how using DE outperforms the state of the art algorithms that were explicitly designed for modeling opponents (in terms average rewards) in two complimentary domains.

Policy search is a class of reinforcement learning algorithms for finding optimal policies in control problems with limited feedback. These methods have shown to be successful in high-dimensional problems, such as robotics control. Though successful, current methods can lead to unsafe policy parameters potentially damaging hardware units. Motivated by such constraints, projection based methods are proposed for safe policies. These methods, however, can only handle convex policy constraints. In this paper, we propose the first safe policy search reinforcement learner capable of operating under non-convex policy constraints. This is achieved by observing, for the first time, a connection between non-convex variational inequalities and policy search problems. We provide two algorithms, i.e., Mann and two-step iteration, to solve the above problems and prove convergence in the non-convex stochastic setting. Finally, we demonstrate the performance of the above algorithms on six benchmark dynamical systems and show that our new method is capable of outperforming previous methods under a variety of settings.

Interactions in multiagent systems are generally more complicated than single agent ones. Game theory provides solutions on how to act in multiagent scenarios; however, it assumes that all agents will act rationally. Moreover, some works also assume the opponent will use a stationary strategy. These assumptions usually do not hold in real world scenarios where agents have limited capacities and may deviate from a perfect rational response. Our goal is still to act optimally in these cases by learning the appropriate response and without any prior policies on how to act. Thus, we focus on the problem when another agent in the environment uses different stationary strategies over time. This will turn the problem into learning in a non-stationary environment, posing a problem for most learning algorithms. This paper introduces DriftER, an algorithm that (1) learns a model of the opponent, (2) uses that to obtain an optimal policy and then (3) determines when it must re-learn due to an opponent strategy change. We provide theoretical results showing that DriftER guarantees to detect switches with high probability. Also, we provide empirical results showing that our approach outperforms state of the art algorithms, in normal form games such as prisoner's dilemma and then in a more realistic scenario, the Power TAC simulator.

The success or failure of any learning algorithm is partially due to the exploration strategy it exerts. However, most exploration strategies assume that the environment is stationary and non-strategic. This work investigates how to design exploration strategies in non-stationary and adversarial environments. Our experimental setting uses a two agents strategic interaction scenario, where the opponent switches between different behavioral patterns. The agent's objective is to learn a model of the opponent's strategy to act optimally, despite non-determinism and stochasticity. Our contribution is twofold. First, we present drift exploration as a strategy for switch detection. Second, we propose a new algorithm called R-max# that reasons and acts in terms of two objectives: 1) to maximize utilities in the short term while learning and 2) eventually explore implicitly looking for opponent behavioral changes. We provide theoretical results showing that R-max# is guaranteed to detect the opponent's switch and learn a new model in terms of finite sample complexity.

Policy search is a class of reinforcement learning algorithms for finding optimal policies in control problems with limited feedback. These methods have been shown to be successful in high-dimensional problems such as robotics control. Though successful, current methods can lead to unsafe policy parameters that potentially could damage hardware units. Motivated by such constraints, we propose projection-based methods for safe policies.These methods, however, can handle only convex policy constraints. In this letter, we propose the first safe policy search reinforcement learner capable of operating under nonconvex policy constraints. This is achieved by observing, for the first time, a connection between nonconvex variational inequalities and policy search problems. We provide two algorithms, Mann and two-step iteration, to solve the above problems and prove convergence in the nonconvex stochastic setting. Finally, we demonstrate the performance of the algorithms on six benchmark dynamical systems and show that our new method is capable of outperforming previous methods under a variety of settings.

Interactions in multiagent systems are generally more complicated than single agent ones. Game theory provides solutions on how to act in multiple agent scenarios; however, it assumes that all agents will act rationally. Moreover, some works also assume the opponent will use a stationary strategy. These assumptions usually do not hold in real world scenarios where agents have limited capacities and may deviate from a perfect rational response. Our goal is still to act optimally in this cases by learning the appropriate response and without any prior policies on how to act. Thus, we focus on the problem when another agent in the environment uses different stationary strategies over time. This paper introduces DriftER, an algorithm that 1) learns a model of the opponent, 2) uses that to obtain an optimal policy and then 3) determines when it must re-learn due to an opponent strategy change. We provide theoretical results showing that DriftER guarantees to detect switches with high probability. Also, we provide empirical results in normal form games and then in a more realistic scenario, the Power TAC simulator.

•Deriving a novel scalable algorithm for lifelong policy search in reinforcement learning.•Acquiring linear convergence rate of our new algorithm•Demonstrating the effectiveness of our technique dynamical systems and learning speed-ups on unobserved tasks.

Policy advice is a transfer learning method where a student agent is able to learn faster via advice from a teacher. However, both this and other reinforcement learning transfer methods have little theoretical analysis. This paper formally defines a setting where multiple teacher agents can provide advice to a student and introduces an algorithm to leverage both autonomous exploration and teacher's advice. Our regret bounds justify the intuition that good teachers help while bad teachers hurt. Using our formalization, we are also able to quantify, for the first time, when negative transfer can occur within such a reinforcement learning setting.

This paper proposes an online transfer framework to capture the interaction among agents and shows that current transfer learning in reinforcement learning is a special case of online transfer. Furthermore, this paper re-characterizes existing agents-teaching-agents methods as online transfer and analyze one such teaching method in three ways. First, the convergence of Q-learning and Sarsa with tabular representation with a finite budget is proven. Second, the convergence of Q-learning and Sarsa with linear function approximation is established. Third, the we show the asymptotic performance cannot be hurt through teaching. Additionally, all theoretical results are empirically validated.

In this paper, we study the computational complexity of solution concepts in the context of coalitional games. Firstly, we distinguish two different kinds of core, the undominated core and excess core, and investigate the difference and relationship between them. Secondly, we thoroughly investigate the computational complexity of undominated core and three farsighted solution concepts---farsighted core, farsighted stable set and largest consistent se}.

The issues of coalition formation have been investigated from many aspects and recently more and more attention has been paid to overlapping coalition formation. The (optimal) coalition structure generation(CSG) problem is one of the essential problems in coalition formation which is an important topic of cooperation in multiagent system. In this paper, we strictly define the coalition structure generation problem of overlapping extensions. Based on these definitions, we systematically prove some computational complexity results for overlapping coalition formation(OCF) games and threshold task games(TTGs). Moreover, dynamic programming and greedy approaches are adopted to solve the CSG for TTG.

Agents increase capabilities and receive more repayment via coalition in multi-agent system. This paper focuses on the improvement of coalition model and coalition formation, and proposes a new coalition model CLAR (coalition model based on learning agent and role), which is based on ARG (agent, role, group) meta model and learning mechanism. It also proposes a two phrase coalition formation mechanism in CLAR model that adopts contract net as its protocol. Finally, the experimental results verify the effect of the role and learning mechanism in predator game, and the effect of two-phrase coalition formation in decreasing and controlling the communication cost.