Seyed Mohammad Asghari

Ian Osband, Zheng Wen, Seyed Mohammad Asghari et al. · stanford

Thompson sampling (TS) is a popular heuristic for action selection, but it requires sampling from a posterior distribution. Unfortunately, this can become computationally intractable in complex environments, such as those modeled using neural networks. Approximate posterior samples can produce effective actions, but only if they reasonably approximate joint predictive distributions of outputs across inputs. Notably, accuracy of marginal predictive distributions does not suffice. Epistemic neural networks (ENNs) are designed to produce accurate joint predictive distributions. We compare a range of ENNs through computational experiments that assess their performance in approximating TS across bandit and reinforcement learning environments. The results indicate that ENNs serve this purpose well and illustrate how the quality of joint predictive distributions drives performance. Further, we demonstrate that the \textit{epinet} -- a small additive network that estimates uncertainty -- matches the performance of large ensembles at orders of magnitude lower computational cost. This enables effective application of TS with computation that scales gracefully to complex environments.

Ian Osband, Seyed Mohammad Asghari, Benjamin Van Roy et al. · stanford

Language models often pre-train on large unsupervised text corpora, then fine-tune on additional task-specific data. However, typical fine-tuning schemes do not prioritize the examples that they tune on. We show that, if you can prioritize informative training data, you can achieve better performance while using fewer labels. To do this we augment a language model with an epinet: a small additional network that helps to estimate model uncertainty and forms an \textit{epistemic neural network} (ENN). ENNs are neural networks that can know what they don't know. Using an epinet to prioritize uncertain data, we can fine-tune BERT on GLUE tasks to the same performance while using 2x less data than training without prioritization. We also investigate performance in synthetic neural network generative models designed to build understanding. In each setting, using an epinet outperforms heuristic active learning schemes.

Vikranth Dwaracherla, Zheng Wen, Ian Osband et al. · deepmind, stanford

In machine learning, an agent needs to estimate uncertainty to efficiently explore and adapt and to make effective decisions. A common approach to uncertainty estimation maintains an ensemble of models. In recent years, several approaches have been proposed for training ensembles, and conflicting views prevail with regards to the importance of various ingredients of these approaches. In this paper, we aim to address the benefits of two ingredients -- prior functions and bootstrapping -- which have come into question. We show that prior functions can significantly improve an ensemble agent's joint predictions across inputs and that bootstrapping affords additional benefits if the signal-to-noise ratio varies across inputs. Our claims are justified by both theoretical and experimental results.

Xiuyuan Lu, Ian Osband, Seyed Mohammad Asghari et al. · stanford

Recent work introduced the epinet as a new approach to uncertainty modeling in deep learning. An epinet is a small neural network added to traditional neural networks, which, together, can produce predictive distributions. In particular, using an epinet can greatly improve the quality of joint predictions across multiple inputs, a measure of how well a neural network knows what it does not know. In this paper, we examine whether epinets can offer similar advantages under distributional shifts. We find that, across ImageNet-A/O/C, epinets generally improve robustness metrics. Moreover, these improvements are more significant than those afforded by even very large ensembles at orders of magnitude lower computational costs. However, these improvements are relatively small compared to the outstanding issues in distributionally-robust deep learning. Epinets may be a useful tool in the toolbox, but they are far from the complete solution.

Seyed Mohammad Asghari, Yi Ouyang, Ashutosh Nayyar

We consider a networked control system consisting of a remote controller and a collection of linear plants, each associated with a local controller. Each local controller directly observes the state of its co-located plant and can inform the remote controller of the plant's state through an unreliable uplink channel. We assume that the downlink channels from the remote controller to local controllers are perfect. The objective of the local controllers and the remote controller is to cooperatively minimize a quadratic performance cost. We provide a dynamic program for this decentralized control problem using the common information approach. Although our problem is not a partially nested problem, we obtain explicit optimal strategies for all controllers. In the optimal strategies, all controllers compute common estimates of the states of the plants based on the common information obtained from the communication network. The remote controller's action is linear in the common state estimates, and the action of each local controller is linear in both the actual state of its co-located plant and the common state estimates. We illustrate our results with numerical experiments using randomly generated models.

Yi Ouyang, Seyed Mohammad Asghari, Ashutosh Nayyar

We consider a decentralized networked control system (DNCS) consisting of a remote controller and a collection of linear plants, each associated with a local controller. Each local controller directly observes the state of its co-located plant and can inform the remote controller of the plant's state through an unreliable uplink channel. The downlink channels from the remote controller to local controllers were assumed to be perfect. The objective of the local controllers and the remote controller is to cooperatively minimize the infinite horizon time average of expected quadratic cost. The finite horizon version of this problem was solved in our prior work [2]. The optimal strategies in the finite horizon case were shown to be characterized by coupled Riccati recursions. In this paper, we show that if the link failure probabilities are below certain critical thresholds, then the coupled Riccati recursions of the finite horizon solution reach a steady state and the corresponding decentralized strategies are optimal. Above these thresholds, we show that no strategy can achieve finite cost. We exploit a connection between our DNCS Riccati recursions and the coupled Riccati recursions of an auxiliary Markov jump linear system to obtain our results. Our main results in Theorems 1 and 2 explicitly identify the critical thresholds for the link failure probabilities and the optimal decentralized control strategies when all link failure probabilities are below their thresholds.

Yi Ouyang, Seyed Mohammad Asghari, Ashutosh Nayyar

We consider a decentralized optimal control problem for a linear plant controlled by two controllers, a local controller and a remote controller. The local controller directly observes the state of the plant and can inform the remote controller of the plant state through a packet-drop channel. We assume that the remote controller is able to send acknowledgments to the local controller to signal the successful receipt of transmitted packets. The objective of the two controllers is to cooperatively minimize a quadratic performance cost. We provide a dynamic program for this decentralized control problem using the common information approach. Although our problem is not a partially nested LQG problem, we obtain explicit optimal strategies for the two controllers. In the optimal strategies, both controllers compute a common estimate of the plant state based on the common information. The remote controller's action is linear in the common estimated state, and the local controller's action is linear in both the actual state and the common estimated state.

Seyed Mohammad Asghari, Ashutosh Nayyar

This paper considers two problems -- a dynamic team problem and a decentralized control problem. The problems we consider do not belong to the known classes of "simpler" dynamic team/decentralized control problems such as partially nested or quadratically invariant problems. However, we show that our problems admit simple solutions under an assumption referred to as the substitutability assumption. Intuitively, substitutability in a team (resp. decentralized control) problem means that the effects of one team member's (resp. controller's) action on the cost function and the information (resp. state dynamics) can be achieved by an action of another member (resp. controller). For the non-partially-nested LQG dynamic team problem, it is shown that under certain conditions linear strategies are optimal. For the non-partially-nested decentralized LQG control problem, the state structure can be exploited to obtain optimal control strategies with recursively update-able sufficient statistics. These results suggest that substitutability can work as a counterpart of the information structure requirements that enable simplification of dynamic teams and decentralized control problems.

Seyed Mohammad Asghari, Ashutosh Nayyar

We consider a decentralized system with multiple controllers and define substitutability of one controller by another in open-loop strategies. We explore the implications of this property on the optimization of closed-loop strategies. In particular, we focus on the decentralized LQG problem with substitutable actions. Even though the problem we formulate does not belong to the known classes of "simpler" decentralized problems such as partially nested or quadratically invariant problems, our results show that, under the substitutability assumption, linear strategies are optimal and we provide a complete state space characterization of optimal strategies. We also identify a family of information structures that all give the same optimal cost as the centralized information structure under the substitutability assumption. Our results suggest that open-loop substitutability can work as a counterpart of the information structure requirements that enable simplification of decentralized control problems.

Seyed Mohammad Asghari, Yi Ouyang, Ashutosh Nayyar

We consider a networked control system (NCS) consisting of two plants, a global plant and a local plant, and two controllers, a global controller and a local controller. The global (resp. local) plant follows discrete-time stochastically switched linear dynamics with a continuous global (resp. local) state and a discrete global (resp. local) mode. We assume that the state and mode of the global plant are observed by both controllers while the state and mode of the local plant are only observed by the local controller. The local controller can inform the global controller of the local plant's state and mode through an unreliable TCP-like communication channel where successful transmissions are acknowledged. The objective of the controllers is to cooperatively minimize a modes-dependent quadratic cost over a finite time horizon. Following the method developed in [1] and [2], we construct a dynamic program based on common information and a decomposition of strategies, and use it to obtain explicit optimal strategies for the controllers. In the optimal strategies, both controllers compute a common estimate of the local plant's state. The global controller's action is linear in the state of the global plant and the common estimated state, and the local controller's action is linear in the actual states of both plants and the common estimated state. Furthermore, the gain matrices for the global controller depend on the global mode and its observation about the local mode, while the gain matrices for the local controller depend on the actual modes of both plants and the global controller's observation about the local mode.

Ian Osband, Zheng Wen, Seyed Mohammad Asghari et al.

Predictive distributions quantify uncertainties ignored by point estimates. This paper introduces The Neural Testbed: an open-source benchmark for controlled and principled evaluation of agents that generate such predictions. Crucially, the testbed assesses agents not only on the quality of their marginal predictions per input, but also on their joint predictions across many inputs. We evaluate a range of agents using a simple neural network data generating process. Our results indicate that some popular Bayesian deep learning agents do not fare well with joint predictions, even when they can produce accurate marginal predictions. We also show that the quality of joint predictions drives performance in downstream decision tasks. We find these results are robust across choice a wide range of generative models, and highlight the practical importance of joint predictions to the community.

Vikranth Dwaracherla, Seyed Mohammad Asghari, Botao Hao et al.

We present evidence of substantial benefit from efficient exploration in gathering human feedback to improve large language models. In our experiments, an agent sequentially generates queries while fitting a reward model to the feedback received. Our best-performing agent generates queries using double Thompson sampling, with uncertainty represented by an epistemic neural network. Our results demonstrate that efficient exploration enables high levels of performance with far fewer queries. Further, both uncertainty estimation and the choice of exploration scheme play critical roles.

Seyed Mohammad Asghari, Chris Chute, Vikranth Dwaracherla et al.

We develop an online learning algorithm that dramatically improves the data efficiency of reinforcement learning from human feedback (RLHF). Our algorithm incrementally updates reward and language models as choice data is received. The reward model is fit to the choice data, while the language model is updated by a variation of reinforce, with reinforcement signals provided by the reward model. Several features enable the efficiency gains: a small affirmative nudge added to each reinforcement signal, an epistemic neural network that models reward uncertainty, and information-directed exploration. With Gemma large language models (LLMs), our algorithm matches the performance of offline RLHF trained on 200K labels using fewer than 20K labels, representing more than a 10x gain in data efficiency. Extrapolating from our results, we expect our algorithm trained on 1M labels to match offline RLHF trained on 1B labels. This represents a 1,000x gain. To our knowledge, these are the first results to demonstrate that such large improvements are possible.

Ian Osband, Zheng Wen, Seyed Mohammad Asghari et al.

Most work on supervised learning research has focused on marginal predictions. In decision problems, joint predictive distributions are essential for good performance. Previous work has developed methods for assessing low-order predictive distributions with inputs sampled i.i.d. from the testing distribution. With low-dimensional inputs, these methods distinguish agents that effectively estimate uncertainty from those that do not. We establish that the predictive distribution order required for such differentiation increases greatly with input dimension, rendering these methods impractical. To accommodate high-dimensional inputs, we introduce \textit{dyadic sampling}, which focuses on predictive distributions associated with random \textit{pairs} of inputs. We demonstrate that this approach efficiently distinguishes agents in high-dimensional examples involving simple logistic regression as well as complex synthetic and empirical data.

Ian Osband, Zheng Wen, Seyed Mohammad Asghari et al.

Intelligence relies on an agent's knowledge of what it does not know. This capability can be assessed based on the quality of joint predictions of labels across multiple inputs. In principle, ensemble-based approaches produce effective joint predictions, but the computational costs of training large ensembles can become prohibitive. We introduce the epinet: an architecture that can supplement any conventional neural network, including large pretrained models, and can be trained with modest incremental computation to estimate uncertainty. With an epinet, conventional neural networks outperform very large ensembles, consisting of hundreds or more particles, with orders of magnitude less computation. The epinet does not fit the traditional framework of Bayesian neural networks. To accommodate development of approaches beyond BNNs, such as the epinet, we introduce the epistemic neural network (ENN) as an interface for models that produce joint predictions.

Seyed Mohammad Asghari, Yi Ouyang, Ashutosh Nayyar

Regret analysis is challenging in Multi-Agent Reinforcement Learning (MARL) primarily due to the dynamical environments and the decentralized information among agents. We attempt to solve this challenge in the context of decentralized learning in multi-agent linear-quadratic (LQ) dynamical systems. We begin with a simple setup consisting of two agents and two dynamically decoupled stochastic linear systems, each system controlled by an agent. The systems are coupled through a quadratic cost function. When both systems' dynamics are unknown and there is no communication among the agents, we show that no learning policy can generate sub-linear in $T$ regret, where $T$ is the time horizon. When only one system's dynamics are unknown and there is one-directional communication from the agent controlling the unknown system to the other agent, we propose a MARL algorithm based on the construction of an auxiliary single-agent LQ problem. The auxiliary single-agent problem in the proposed MARL algorithm serves as an implicit coordination mechanism among the two learning agents. This allows the agents to achieve a regret within $O(\sqrt{T})$ of the regret of the auxiliary single-agent problem. Consequently, using existing results for single-agent LQ regret, our algorithm provides a $\tilde{O}(\sqrt{T})$ regret bound. (Here $\tilde{O}(\cdot)$ hides constants and logarithmic factors). Our numerical experiments indicate that this bound is matched in practice. From the two-agent problem, we extend our results to multi-agent LQ systems with certain communication patterns.

Navid Naderializadeh, Seyed Mohammad Asghari

We consider a system comprising a file library and a network with a server and multiple users equipped with cache memories. The system operates in two phases: a prefetching phase, where users load their caches with parts of contents from the library, and a delivery phase, where users request files from the library and the server needs to send the uncached parts of the requested files to the users. For the case where the users' caches are arbitrarily loaded, we propose an algorithm based on deep reinforcement learning to minimize the delay of delivering requested contents to the users in the delivery phase. Simulation results demonstrate that our proposed deep reinforcement learning agent learns a coded delivery strategy for sending the requests to the users, which slightly outperforms the state-of-the-art performance in terms of delivery delay, while drastically reducing the computational complexity.