New Zealand

Takeshi Shibuya and Tomoki Hamagami (January 14th 2011). Complex-Valued Reinforcement Learning: a Context-Based Approach for POMDPs, Advances in Reinforcement Learning, Abdelhamid Mellouk, IntechOpen, DOI: 10.5772/13295. 

Le Tien Dung, Takashi Komeda and Motoki Takagi (January 1st 2009). Solving POMDPs with Automatic Discovery of Subgoals, Theory and Novel Applications of Machine Learning, Meng Joo Er and Yi Zhou, IntechOpen, DOI: 10.5772/6685. 

Online advertising campaigns are typically launched for a customer across multiple touch points (scenarios) before the conversion of his final purchase. To maximize the advertisers’ revenue, it requires the platform to develop its advertising strategy based on the consumers’ behavioral trajectories in the previous scenarios. Meanwhile, it is also critical to maintain the interpretability of the models on the conversion rate; however, modern reinforcement learning based solutions fail to do so due to their black-box modeling on the consumer intents. In this paper, we model consumer’s purchase intention as a latent variable, and formulate the advertising problem as a partially observed Markov Decision Process (POMDP). We incorporate the expectation maximization (EM) algorithms for solving the optimal POMDP. Our extensive experiments based on large-scale real-world data demonstrate that our method provides superior performance over several baselines. Apart from the improved advertising performance, our method is able to offer interpretation on the attribution of the conversion.


AAMAS 2020

Sequential Advertising Agent with Interpretable User Hidden Intents

user hidden states

online-advertising

pomdp

AAMAS is the leading scientific conference for research in autonomous agents and multi-agent systems. The AAMAS conference series was initiated in 2002 as the merging of three respected scientific meetings: the International Conference on Multi-Agent Systems (ICMAS), the International Workshop on Agent Theories, Architectures, and Languages (ATAL), and the International Conference on Autonomous Agents (AA). The aim of the joint conference is to provide a single, high-profile, internationally-respected archival forum for scientific research in the theory and practice of autonomous agents and multi-agent systems.

Browse keynotes, discussions, panels and over 300 presentations.


AAMAS is the leading scientific conference for research in autonomous agents and multi-agent systems. The AAMAS conference series was initiated in 2002 as the merging of three respected scientific meetings: the International Conference on Multi-Agent Systems (ICMAS), the International Workshop on Agent Theories, Architectures, and Languages (ATAL), and the International Conference on Autonomous Agents (AA). 

technical paper

Aggregation of Support-Relations of BAFs

In this work we present a novel approach for transfer-guided exploration in reinforcement learning that is inspired by the human tendency to leverage experiences from similar encounters in the past while navigating a new task. Given an optimal policy in a related task-environment, we show that its bisimulation distance from the current task-environment gives a lower bound on the optimal advantage of state-action pairs in the current task-environment. Transfer-guided Exploration (ExTra) samples actions from a Softmax distribution over these lower bounds. In this way, actions with potentially higher optimum advantage are sampled more frequently. In our experiments on gridworld environments, we demonstrate that given access to an optimal policy in a related task-environment, ExTra can outperform popular domain-specific exploration strategies viz. epsilon greedy, Model-Based Interval Estimation - Exploration Based (MBIE-EB), Pursuit and Boltzmann in terms of sample complexity and rate of convergence. We further show that ExTra is robust to choices of source task and shows a graceful degradation of performance as the dissimilarity of the source task increases. We also demonstrate that ExTra, when used alongside traditional exploration algorithms, improves their rate of convergence. Thus it is capable of complementing the efficacy of traditional exploration algorithms.


ExTra: Transfer-guided Exploration

Neural networks that output the parameters of a diagonal Gaussian distribution are widely used in reinforcement learning tasks with continuous action spaces. They have had considerable success in single-agent domains and even in some multi-agent tasks. However, general multi-agent tasks often require mixed strategies whose distributions cannot be well approximated by Gaussians or their mixtures. This paper proposes an alternative for policy representation based on normalizing flows. This approach allows for greater flexibility in action distribution representation beyond mixture models. We demonstrate their advantage over standard methods on a set of imitation learning tasks modeling human driving behaviors in the presence of other drivers.


Normalizing Flow Model for Policy Representation in Continuous Action Multi-agent Systems

Multi-agent coordination is prevalent in many real-world applications. However, such coordination is challenging due to its combinatorial nature. An important observation in this regard is that agents in the real world often only directly affect a limited set of neighbouring agents. Leveraging such loose couplings among agents is key to making coordination in multi-agent systems feasible.
In this work, we focus on learning to coordinate.
Specifically, we consider the multi-agent multi-armed bandit framework, in which fully cooperative loosely-coupled agents must learn to coordinate their decisions to optimize a common objective. 
We propose multi-agent Thompson sampling (MATS), a new Bayesian exploration-exploitation algorithm that leverages loose couplings. We empirically show that MATS outperforms the state-of-the-art algorithm, MAUCE, on two synthetic benchmarks with Bernoulli-distributed rewards. Our results show that MATS improves significantly upon state-of-the-art coordination methods in terms of performance, demonstrating the value of using MATS in applications with sparse neighbourhood structures.


Thompson Sampling for Factored Multi-Agent Bandits

Despite the success of deep reinforcement learning in a variety type of games such as Board games, RTS, FPS, and MOBA games, sports games (SPG) like basketball have been seldom studied. Basketball is one of the most popular and challenging sports games due to its long-time horizon, sparse rewards, complex game rules, and multiple roles with different capabilities. Although these problems could be partially alleviated by common methods like hierarchical reinforcement learning through a decomposition of the whole game into several subtasks based on game rules (such as attack, defense), these methods tend to ignore the strong correlations between these subtasks and could have difficulty in generating reasonable poli- cies across the whole basketball match. Besides, the existence of multiple agents adds extra challenges to such game. In this work, we propose an integrated curriculum training approach (ICTA) which is composed of two parts. The first part is for handling the correlated subtasks from the perspective of a single player, which contains several weighted cascading curriculum learners that can smoothly unify the base curriculum training of corresponding sub- tasks together using a Q-value backup mechanism with a weight factor. The second part is for enhancing the cooperation ability of the basketball team, which is a curriculum switcher that focuses on learning the switch of the cooperative curriculum within one team by taking over collaborative actions such as passing from single- player’s action spaces. Our method is then applied to a commercial online basketball game named Fever Basketball (FB). Results show that ICTA significantly outperforms the built-in AI and reaches up to around 70% win-rate than online human players during a 300-day evaluation period.


Mastering Basketball with Deep Reinforcement Learning: An Integrated Curriculum Training Approach

We consider Deep Reinforcement Learning (DRL) approaches to devise mapless navigation strategies for mobile platforms. We propose a Genetic Deep Reinforcement Learning (GDRL) method that combines Genetic Algorithms (GA) with discrete and continuous action space DRL approaches. The goal of GDRL is to reduce the sensitivity of DRL approaches to their hyper-parameter tuning and to provide robust exploration strategies. We evaluate GDRL in combination with Rainbow and Proximal Policy Optimization (PPO) in two navigation scenarios: i) a wheeled robot avoiding obstacles in an indoor environment and ii) a water drone that must reach a predefined location in presence of waves. Our empirical evaluation demonstrates that GDRL outperforms state-of-the-art DRL and GA methods as well as a previous hybrid approach.


Genetic Deep Reinforcement Learning for Mapless Navigation

Multi-agent path finding (MAPF), naturally exists in applications like picking-up and dropping-off parcels by automated guided vehicles (AGVs) in the warehouse. Existing algorithms, like conflict-based search (CBS), windowed hierarchical cooperative A* (WHCA), and other A* variants, are widely used to find the shortest paths in different manners. However, in real-world environments, MAPF cases are dynamically generated and need to be solved in real time. In this work, a decentralized multi-agent reinforcement learning (MARL) framework with multi-step ahead tree search (MATS) strategy is proposed to make efficient decisions. Through performing experiments on a 30*30 grid world and a real-world warehouse case, our proposed MARL policy is proved to be capable of: 1) scaling to a large number of agents in real-world environment with online response time within acceptable levels; 2) outperforming existing algorithms with shorter path length and solution time, as the number of agents increases.


Learning to Cooperate: Application of Deep Reinforcement Learning for Online AGV Path Finding

In this talk, we study a variant of hedonic games, called Seat Arrangement. 
The model is defined by a bijection from agents with preferences to vertices in a graph.
The utility of an agent depends on the neighbors in the graph. 
In this talk, we study the price of stability and fairness in Seat Arrangement, and the computational complexity and the parameterized complexity of finding certain desirable seat arrangements, say Maximum Welfare Arrangement, Maximin Utility Arrangement, Stable Arrangement, and Envy-free Arrangement.


Hedonic Seat Arrangement Problems

In order to be useful in the real world, an AI agent needs to plan and act in the presence of other agents, who may be helpful or disruptive. In this paper, we consider the problem where an autonomous agent needs to act in a manner that clarifies its objectives to cooperative agents while simultaneously preventing adversarial agents from inferring those objectives. We call it Mixed-Observer Controlled Observability Planning Problem (MO-COPP). We develop two new solution approaches: one provides an optimal solution to the problem given a fixed time horizon by using an integer programming solver, the other provides a satisficing solution using heuristic-guided forward search to achieve prespecified amount of obfuscation and legibility for adversarial and cooperative agents respectively. 


Signaling Friends and Head-Faking Enemies Simultaneously: Balancing Goal Obfuscation and Goal Legibility

In many multi-agent interactions in the real world, agents receive payoffs over multiple distinct criteria; i.e. the payoffs are multi-objective in nature. However, the same multi-objective payoff vector may lead to different utilities for each participant. Therefore, it is essential for an agent to learn about the behaviour of other agents in the system. We investigate here the effects of opponent modelling on multi-objective multi-agent interactions with non-linear utilities. Specifically, we consider multi-objective normal form games (MONFGs) with non-linear utility functions under the scalarised expected returns optimisation criterion. We contribute a novel actor-critic formulation to allow reinforcement learning of mixed strategies in this setting, along with an extension that incorporates opponent policy reconstruction using conditional action frequencies. Our empirical results demonstrate that opponent modelling can drastically alter the learning dynamics in this setting.


Downloads

Next from AAMAS 2020

Aggregation of Support-Relations of BAFs

Similar lecture

Goal-driven Adaptation of MOISE Organizations for Workflow Enactment

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from AAMAS 2020

Aggregation of Support-Relations of BAFs

Similar lecture

Goal-driven Adaptation of MOISE Organizations for Workflow Enactment

Downloads