Autonomous exploration has many important applications. However, classic information gain-based or frontier-based exploration only relies on the robot current state to determine the immediate exploration goal, which lacks the capability of predicting the value of future states and thus leads to inefficient exploration decisions. This paper presents a method to learn how "good" states are, measured by the state value function, to provide a guidance for robot exploration in real-world challenging environments. We formulate our work as an off-policy evaluation (OPE) problem for robot exploration (OPERE). It consists of offline Monte-Carlo training on real-world data and performs Temporal Difference (TD) online adaptation to optimize the trained value estimator. We also design an intrinsic reward function based on sensor information coverage to enable the robot to gain more information with sparse extrinsic rewards. Results show that our method enables the robot to predict the value of future states so as to better guide robot exploration. The proposed algorithm achieves better prediction and exploration performance compared with the state-of-the-arts. To the best of our knowledge, this work for the first time demonstrates value function prediction on real-world dataset for robot exploration in challenging subterranean and urban environments. More details and demo videos can be found at https://jeffreyyh.github.io/opere/.

In a multi-agent environment, In order to overcome and alleviate the non-stationarity of the multi-agent environment, the mainstream method is to adopt the framework of Centralized Training Decentralized Execution (CTDE). This thesis is based on the framework of CTDE, and studies the cooperative decision-making of multi-agent based on the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm for multi-agent proximal policy optimization. In order to alleviate the non-stationarity of the multi-agent environment, a multi-agent communication mechanism based on weight scheduling and attention module is introduced. Different agents can alleviate the non-stationarity caused by local observations through information exchange between agents, assisting in the collaborative decision-making of agents. The specific method is to introduce a communication module in the policy network part. The communication module is composed of a weight generator, a weight scheduler, a message encoder, a message pool and an attention module. Among them, the weight generator and weight scheduler will generate weights as the selection basis for communication, the message encoder is used to compress and encode communication information, the message pool is used to store communication messages, and the attention module realizes the interactive processing of the agent's own information and communication information. This thesis proposes a Multi-Agent Communication and Global Information Optimization Proximal Policy Optimization(MCGOPPO)algorithm, and conducted experiments in the SMAC and the MPE. The experimental results show that the improvement has achieved certain effects, which can better alleviate the non-stationarity of the multi-agent environment, and improve the collaborative decision-making ability among the agents.

A new trend in deep learning, represented by Mutual Information Neural Estimation (MINE) and Information Noise Contrast Estimation (InfoNCE), is emerging. In this trend, similarity functions and Estimated Mutual Information (EMI) are used as learning and objective functions. Coincidentally, EMI is essentially the same as Semantic Mutual Information (SeMI) proposed by the author 30 years ago. This paper first reviews the evolutionary histories of semantic information measures and learning functions. Then, it briefly introduces the author's semantic information G theory with the rate-fidelity function R(G) (G denotes SeMI, and R(G) extends R(D)) and its applications to multi-label learning, the maximum Mutual Information (MI) classification, and mixture models. Then it discusses how we should understand the relationship between SeMI and Shan-non's MI, two generalized entropies (fuzzy entropy and coverage entropy), Autoencoders, Gibbs distributions, and partition functions from the perspective of the R(G) function or the G theory. An important conclusion is that mixture models and Restricted Boltzmann Machines converge because SeMI is maximized, and Shannon's MI is minimized, making information efficiency G/R close to 1. A potential opportunity is to simplify deep learning by using Gaussian channel mixture models for pre-training deep neural networks' latent layers without considering gradients. It also discusses how the SeMI measure is used as the reward function (reflecting purposiveness) for reinforcement learning. The G theory helps interpret deep learning but is far from enough. Combining semantic information theory and deep learning will accelerate their development.

Many recent developments in large language models focus on prompting them to perform specific tasks. One effective prompting method is in-context learning, where the model performs a (possibly new) generation/prediction task given one (or more) examples. Past work has shown that the choice of examples can make a large impact on task performance. However, finding good examples is not straightforward since the definition of a representative group of examples can vary greatly depending on the task. While there are many existing methods for selecting in-context examples, they generally score examples independently, ignoring the dependency between them and the order in which they are provided to the large language model. In this work, we propose Retrieval for In-Context Learning (RetICL), a learnable method for modeling and optimally selecting examples sequentially for in-context learning. We frame the problem of sequential example selection as a Markov decision process, design an example retriever model using an LSTM, and train it using proximal policy optimization (PPO). We validate RetICL on math problem solving datasets and show that it outperforms both heuristic and learnable baselines, and achieves state-of-the-art accuracy on the TabMWP dataset. We also use case studies to show that RetICL implicitly learns representations of math problem solving strategies.

The Stochastic Gradient Langevin Dynamics (SGLD) are popularly used to approximate Bayesian posterior distributions in statistical learning procedures with large-scale data. As opposed to many usual Markov chain Monte Carlo (MCMC) algorithms, SGLD is not stationary with respect to the posterior distribution; two sources of error appear: The first error is introduced by an Euler--Maruyama discretisation of a Langevin diffusion process, the second error comes from the data subsampling that enables its use in large-scale data settings. In this work, we consider an idealised version of SGLD to analyse the method's pure subsampling error that we then see as a best-case error for diffusion-based subsampling MCMC methods. Indeed, we introduce and study the Stochastic Gradient Langevin Diffusion (SGLDiff), a continuous-time Markov process that follows the Langevin diffusion corresponding to a data subset and switches this data subset after exponential waiting times. There, we show that the Wasserstein distance between the posterior and the limiting distribution of SGLDiff is bounded above by a fractional power of the mean waiting time. Importantly, this fractional power does not depend on the dimension of the state space. We bring our results into context with other analyses of SGLD.

Game theoretic methods have become popular for planning and prediction in situations involving rich multi-agent interactions. However, these methods often assume the existence of a single local Nash equilibria and are hence unable to handle uncertainty in the intentions of different agents. While maximum entropy (MaxEnt) dynamic games try to address this issue, practical approaches solve for MaxEnt Nash equilibria using linear-quadratic approximations which are restricted to unimodal responses and unsuitable for scenarios with multiple local Nash equilibria. By reformulating the problem as a POMDP, we propose MPOGames, a method for efficiently solving MaxEnt dynamic games that captures the interactions between local Nash equilibria. We show the importance of uncertainty-aware game theoretic methods via a two-agent merge case study. Finally, we prove the real-time capabilities of our approach with hardware experiments on a 1/10th scale car platform.

This paper provides a simulated laboratory for making use of Reinforcement Learning (RL) for chemical discovery. Since RL is fairly data intensive, training agents `on-the-fly' by taking actions in the real world is infeasible and possibly dangerous. Moreover, chemical processing and discovery involves challenges which are not commonly found in RL benchmarks and therefore offer a rich space to work in. We introduce a set of highly customizable and open-source RL environments, ChemGymRL, based on the standard Open AI Gym template. ChemGymRL supports a series of interconnected virtual chemical benches where RL agents can operate and train. The paper introduces and details each of these benches using well-known chemical reactions as illustrative examples, and trains a set of standard RL algorithms in each of these benches. Finally, discussion and comparison of the performances of several standard RL methods are provided in addition to a list of directions for future work as a vision for the further development and usage of ChemGymRL.

We study the problem of Inverse Reinforcement Learning (IRL) with an average-reward criterion. The goal is to recover an unknown policy and a reward function when the agent only has samples of states and actions from an experienced agent. Previous IRL methods assume that the expert is trained in a discounted environment, and the discount factor is known. This work alleviates this assumption by proposing an average-reward framework with efficient learning algorithms. We develop novel stochastic first-order methods to solve the IRL problem under the average-reward setting, which requires solving an Average-reward Markov Decision Process (AMDP) as a subproblem. To solve the subproblem, we develop a Stochastic Policy Mirror Descent (SPMD) method under general state and action spaces that needs $\mathcal{{O}}(1/\varepsilon)$ steps of gradient computation. Equipped with SPMD, we propose the Inverse Policy Mirror Descent (IPMD) method for solving the IRL problem with a $\mathcal{O}(1/\varepsilon^2)$ complexity. To the best of our knowledge, the aforementioned complexity results are new in IRL. Finally, we corroborate our analysis with numerical experiments using the MuJoCo benchmark and additional control tasks.

In recent years, large language models (LMs) have achieved remarkable progress across various natural language processing tasks. As pre-training and fine-tuning are costly and might negatively impact model performance, it is desired to efficiently adapt an existing model to different conditions such as styles, sentiments or narratives, when facing different audiences or scenarios. However, efficient adaptation of a language model to diverse conditions remains an open challenge. This work is inspired by the observation that text conditions are often associated with selection of certain words in a context. Therefore we introduce LM-Switch, a theoretically grounded, lightweight and simple method for generative language model conditioning. We begin by investigating the effect of conditions in Hidden Markov Models (HMMs), and establish a theoretical connection with language model. Our finding suggests that condition shifts in HMMs are associated with linear transformations in word embeddings. LM-Switch is then designed to deploy a learnable linear factor in the word embedding space for language model conditioning. We show that LM-Switch can model diverse tasks, and achieves comparable or better performance compared with state-of-the-art baselines in LM detoxification and generation control, despite requiring no more than 1% of parameters compared with baselines and little extra time overhead compared with base LMs. It is also able to learn from as few as a few sentences or one document. Moreover, a learned LM-Switch can be transferred to other LMs of different sizes, achieving a detoxification performance similar to the best baseline. We will make our code available to the research community following publication.

We present an approach for autonomous sensor control for information gathering under partially observable, dynamic and sparsely sampled environments that maximizes information about entities present in that space. We describe our approach for the task of Radio-Frequency (RF) spectrum monitoring, where the goal is to search for and track unknown, dynamic signals in the environment. To this end, we extend the Deep Anticipatory Network (DAN) Reinforcement Learning (RL) framework by (1) improving exploration in sparse, non-stationary environments using a novel information gain reward, and (2) scaling up the control space and enabling the monitoring of complex, dynamic activity patterns using hybrid convolutional-recurrent neural layers. We also extend this problem to situations in which sampling from the intended RF spectrum/field is limited and propose a model-based version of the original RL algorithm that fine-tunes the controller via a model that is iteratively improved from the limited field sampling. Results in simulated RF environments of differing complexity show that our system outperforms the standard DAN architecture and is more flexible and robust than baseline expert-designed agents. We also show that it is adaptable to non-stationary emission environments.