Dylan R. Ashley

Miroslav Štrupl, Francesco Faccio, Dylan R. Ashley et al.

Upside-Down Reinforcement Learning (UDRL) is an approach for solving RL problems that does not require value functions and uses only supervised learning, where the targets for given inputs in a dataset do not change over time. Ghosh et al. proved that Goal-Conditional Supervised Learning (GCSL) -- which can be viewed as a simplified version of UDRL -- optimizes a lower bound on goal-reaching performance. This raises expectations that such algorithms may enjoy guaranteed convergence to the optimal policy in arbitrary environments, similar to certain well-known traditional RL algorithms. Here we show that for a specific episodic UDRL algorithm (eUDRL, including GCSL), this is not the case, and give the causes of this limitation. To do so, we first introduce a helpful rewrite of eUDRL as a recursive policy update. This formulation helps to disprove its convergence to the optimal policy for a wide class of stochastic environments. Finally, we provide a concrete example of a very simple environment where eUDRL diverges. Since the primary aim of this paper is to present a negative result, and the best counterexamples are the simplest ones, we restrict all discussions to finite (discrete) environments, ignoring issues of function approximation and limited sample size.

Dylan R. Ashley, Vincent Herrmann, Zachary Friggstad et al.

The act of telling stories is a fundamental part of what it means to be human. This work introduces the concept of narrative information, which we define to be the overlap in information space between a story and the items that compose the story. Using contrastive learning methods, we show how modern artificial neural networks can be leveraged to distill stories and extract a representation of the narrative information. We then demonstrate how evolutionary algorithms can leverage this to extract a set of narrative templates and how these templates -- in tandem with a novel curve-fitting algorithm we introduce -- can reorder music albums to automatically induce stories in them. In the process of doing so, we give strong statistical evidence that these narrative information templates are present in existing albums. While we experiment only with music albums here, the premises of our work extend to any form of (largely) independent media.

Dylan R. Ashley, Gaël Le Lan, Changsheng Zhao et al.

Large language models (LLMs) face a fundamental trade-off between computational efficiency (e.g., number of parameters) and output quality, especially when deployed on computationally limited devices such as phones or laptops. One way to address this challenge is by following the example of humans and have models ask for help when they believe they are incapable of solving a problem on their own; we can overcome this trade-off by allowing smaller models to respond to queries when they believe they can provide good responses, and deferring to larger models when they do not believe they can. To this end, in this paper, we investigate the viability of Predict-Answer/Act (PA) and Reason-Predict-Reason-Answer/Act (RPRA) paradigms where models predict -- prior to responding -- how an LLM judge would score their output. We evaluate three approaches: zero-shot prediction, prediction using an in-context report card, and supervised fine-tuning. Our results show that larger models (particularly reasoning models) perform well when predicting generic LLM judges zero-shot, while smaller models can reliably predict such judges well after being fine-tuned or provided with an in-context report card. Altogether, both approaches can substantially improve the prediction accuracy of smaller models, with report cards and fine-tuning achieving mean improvements of up to 55% and 52% across datasets, respectively. These findings suggest that models can learn to predict their own performance limitations, paving the way for more efficient and self-aware AI systems.

Wenyi Wang, Hisham A. Alyahya, Dylan R. Ashley et al.

Language-based agentic systems have shown great promise in recent years, transitioning from solving small-scale research problems to being deployed in challenging real-world tasks. However, optimizing these systems often requires substantial manual labor. Recent studies have demonstrated that these systems can be represented as computational graphs, enabling automatic optimization. Despite these advancements, most current efforts in Graph-based Agentic System Optimization (GASO) fail to properly assign feedback to the system's components given feedback on the system's output. To address this challenge, we formalize the concept of semantic backpropagation with semantic gradients -- a generalization that aligns several key optimization techniques, including reverse-mode automatic differentiation and the more recent TextGrad by exploiting the relationship among nodes with a common successor. This serves as a method for computing directional information about how changes to each component of an agentic system might improve the system's output. To use these gradients, we propose a method called semantic gradient descent which enables us to solve GASO effectively. Our results on both BIG-Bench Hard and GSM8K show that our approach outperforms existing state-of-the-art methods for solving GASO problems. A detailed ablation study on the LIAR dataset demonstrates the parsimonious nature of our method. A full copy of our implementation is publicly available at https://github.com/HishamAlyahya/semantic_backprop

Yanning Dai, Yuhui Wang, Dylan R. Ashley et al.

Morphology-control co-design concerns the coupled optimization of an agent's body structure and control policy. This problem exhibits a bi-level structure, where the control dynamically adapts to the morphology to maximize performance. Existing methods typically neglect the control's adaptation dynamics by adopting a single-level formulation that treats the control policy as fixed when optimizing morphology. This can lead to inefficient optimization, as morphology updates may be misaligned with control adaptation. In this paper, we revisit the co-design problem from a game-theoretic perspective, modeling the intrinsic coupling between morphology and control as a novel variant of a Stackelberg game. We propose Stackelberg Proximal Policy Optimization (Stackelberg PPO), which explicitly incorporates the control's adaptation dynamics into morphology optimization. By modeling this intrinsic coupling, our method aligns morphology updates with control adaptation, thereby stabilizing training and improving learning efficiency. Experiments across diverse co-design tasks demonstrate that Stackelberg PPO outperforms standard PPO in both stability and final performance, opening the way for dramatically more efficient robotics designs.

Jacopo Di Ventura, Dylan R. Ashley, Vincent Herrmann et al.

Upside Down Reinforcement Learning (UDRL) is a promising framework for solving reinforcement learning problems which focuses on learning command-conditioned policies. In this work, we extend UDRL to the task of learning a command-conditioned generator of deep neural network policies. We accomplish this using Hypernetworks - a variant of Fast Weight Programmers, which learn to decode input commands representing a desired expected return into command-specific weight matrices. Our method, dubbed Upside Down Reinforcement Learning with Policy Generators (UDRLPG), streamlines comparable techniques by removing the need for an evaluator or critic to update the weights of the generator. To counteract the increased variance in last returns caused by not having an evaluator, we decouple the sampling probability of the buffer from the absolute number of policies in it, which, together with a simple weighting strategy, improves the empirical convergence of the algorithm. Compared with existing algorithms, UDRLPG achieves competitive performance and high returns, sometimes outperforming more complex architectures. Our experiments show that a trained generator can generalize to create policies that achieve unseen returns zero-shot. The proposed method appears to be effective in mitigating some of the challenges associated with learning highly multimodal functions. Altogether, we believe that UDRLPG represents a promising step forward in achieving greater empirical sample efficiency in RL. A full implementation of UDRLPG is publicly available at https://github.com/JacopoD/udrlpg_

Vincent Herrmann, Dylan R. Ashley, Jürgen Schmidhuber

Album sequencing is a critical part of the album production process. Recently, a data-driven approach was proposed that sequences general collections of independent media by extracting the narrative essence of the items in the collections. While this approach implies an album sequencing technique, it is not widely accessible to a less technical audience, requiring advanced knowledge of machine learning techniques to use. To address this, we introduce a new user-friendly web-based tool that allows a less technical audience to upload music tracks, execute this technique in one click, and subsequently presents the result in a clean visualization to the user. To both increase the number of templates available to the user and address shortcomings of previous work, we also introduce a new direct transformer-based album sequencing method. We find that our more direct method outperforms a random baseline but does not reach the same performance as the narrative essence approach. Both methods are included in our web-based user interface, and this -- alongside a full copy of our implementation -- is publicly available at https://github.com/dylanashley/automatic-album-sequencing

Dylan R. Ashley, Vincent Herrmann, Zachary Friggstad et al.

We look at how machine learning techniques that derive properties of items in a collection of independent media can be used to automatically embed stories into such collections. To do so, we use models that extract the tempo of songs to make a music playlist follow a narrative arc. Our work specifies an open-source tool that uses pre-trained neural network models to extract the global tempo of a set of raw audio files and applies these measures to create a narrative-following playlist. This tool is available at https://github.com/dylanashley/playlist-story-builder/releases/tag/v1.0.0

Miroslav Štrupl, Oleg Szehr, Francesco Faccio et al.

This article provides a rigorous analysis of convergence and stability of Episodic Upside-Down Reinforcement Learning, Goal-Conditioned Supervised Learning and Online Decision Transformers. These algorithms performed competitively across various benchmarks, from games to robotic tasks, but their theoretical understanding is limited to specific environmental conditions. This work initiates a theoretical foundation for algorithms that build on the broad paradigm of approaching reinforcement learning through supervised learning or sequence modeling. At the core of this investigation lies the analysis of conditions on the underlying environment, under which the algorithms can identify optimal solutions. We also assess whether emerging solutions remain stable in situations where the environment is subject to tiny levels of noise. Specifically, we study the continuity and asymptotic convergence of command-conditioned policies, values and the goal-reaching objective depending on the transition kernel of the underlying Markov Decision Process. We demonstrate that near-optimal behavior is achieved if the transition kernel is located in a sufficiently small neighborhood of a deterministic kernel. The mentioned quantities are continuous (with respect to a specific topology) at deterministic kernels, both asymptotically and after a finite number of learning cycles. The developed methods allow us to present the first explicit estimates on the convergence and stability of policies and values in terms of the underlying transition kernels. On the theoretical side we introduce a number of new concepts to reinforcement learning, like working in segment spaces, studying continuity in quotient topologies and the application of the fixed-point theory of dynamical systems. The theoretical study is accompanied by a detailed investigation of example environments and numerical experiments.

Yuhui Wang, Qingyuan Wu, Dylan R. Ashley et al.

The Value Iteration Network (VIN) is an end-to-end differentiable neural network architecture for planning. It exhibits strong generalization to unseen domains by incorporating a differentiable planning module that operates on a latent Markov Decision Process (MDP). However, VINs struggle to scale to long-term and large-scale planning tasks, such as navigating a 100x100 maze -- a task that typically requires thousands of planning steps to solve. We observe that this deficiency is due to two issues: the representation capacity of the latent MDP and the planning module's depth. We address these by augmenting the latent MDP with a dynamic transition kernel, dramatically improving its representational capacity, and, to mitigate the vanishing gradient problem, introduce an "adaptive highway loss" that constructs skip connections to improve gradient flow. We evaluate our method on 2D/3D maze navigation environments, continuous control, and the real-world Lunar rover navigation task. We find that our new method, named Dynamic Transition VIN (DT-VIN), scales to 5000 layers and solves challenging versions of the above tasks. Altogether, we believe that DT-VIN represents a concrete step forward in performing long-term large-scale planning in complex environments.

Mohannad Alhakami, Dylan R. Ashley, Joel Dunham et al.

Advanced machine learning algorithms require platforms that are extremely robust and equipped with rich sensory feedback to handle extensive trial-and-error learning without relying on strong inductive biases. Traditional robotic designs, while well-suited for their specific use cases, are often fragile when used with these algorithms. To address this gap -- and inspired by the vision of enabling curiosity-driven baby robots -- we present a novel robotic limb designed from scratch. Our design has a hybrid soft-hard structure, high redundancy with rich non-contact sensors (exclusively cameras), and easily replaceable failure points. Proof-of-concept experiments using two contemporary reinforcement learning algorithms on a physical prototype demonstrate that our design is able to succeed in a simple target-finding task even under simulated sensor failures, all with minimal human oversight during extended learning periods. We believe this design represents a concrete step toward more tailored robotic designs for achieving general-purpose, generally intelligent robots.

Mingchen Zhuge, Haozhe Liu, Francesco Faccio et al.

Both Minsky's "society of mind" and Schmidhuber's "learning to think" inspire diverse societies of large multimodal neural networks (NNs) that solve problems by interviewing each other in a "mindstorm." Recent implementations of NN-based societies of minds consist of large language models (LLMs) and other NN-based experts communicating through a natural language interface. In doing so, they overcome the limitations of single LLMs, improving multimodal zero-shot reasoning. In these natural language-based societies of mind (NLSOMs), new agents -- all communicating through the same universal symbolic language -- are easily added in a modular fashion. To demonstrate the power of NLSOMs, we assemble and experiment with several of them (having up to 129 members), leveraging mindstorms in them to solve some practical AI tasks: visual question answering, image captioning, text-to-image synthesis, 3D generation, egocentric retrieval, embodied AI, and general language-based task solving. We view this as a starting point towards much larger NLSOMs with billions of agents-some of which may be humans. And with this emergence of great societies of heterogeneous minds, many new research questions have suddenly become paramount to the future of artificial intelligence. What should be the social structure of an NLSOM? What would be the (dis)advantages of having a monarchical rather than a democratic structure? How can principles of NN economies be used to maximize the total reward of a reinforcement learning NLSOM? In this work, we identify, discuss, and try to answer some of these questions.

Kai Arulkumaran, Dylan R. Ashley, Jürgen Schmidhuber et al.

Upside down reinforcement learning (UDRL) flips the conventional use of the return in the objective function in RL upside down, by taking returns as input and predicting actions. UDRL is based purely on supervised learning, and bypasses some prominent issues in RL: bootstrapping, off-policy corrections, and discount factors. While previous work with UDRL demonstrated it in a traditional online RL setting, here we show that this single algorithm can also work in the imitation learning and offline RL settings, be extended to the goal-conditioned RL setting, and even the meta-RL setting. With a general agent architecture, a single UDRL agent can learn across all paradigms.

Dylan R. Ashley, Kai Arulkumaran, Jürgen Schmidhuber et al.

Lately, there has been a resurgence of interest in using supervised learning to solve reinforcement learning problems. Recent work in this area has largely focused on learning command-conditioned policies. We investigate the potential of one such method -- upside-down reinforcement learning -- to work with commands that specify a desired relationship between some scalar value and the observed return. We show that upside-down reinforcement learning can learn to carry out such commands online in a tabular bandit setting and in CartPole with non-linear function approximation. By doing so, we demonstrate the power of this family of methods and open the way for their practical use under more complicated command structures.

Miroslav Štrupl, Francesco Faccio, Dylan R. Ashley et al.

Reward-Weighted Regression (RWR) belongs to a family of widely known iterative Reinforcement Learning algorithms based on the Expectation-Maximization framework. In this family, learning at each iteration consists of sampling a batch of trajectories using the current policy and fitting a new policy to maximize a return-weighted log-likelihood of actions. Although RWR is known to yield monotonic improvement of the policy under certain circumstances, whether and under which conditions RWR converges to the optimal policy have remained open questions. In this paper, we provide for the first time a proof that RWR converges to a global optimum when no function approximation is used, in a general compact setting. Furthermore, for the simpler case with finite state and action spaces we prove R-linear convergence of the state-value function to the optimum.

Dylan R. Ashley, Sina Ghiassian, Richard S. Sutton

Catastrophic forgetting remains a severe hindrance to the broad application of artificial neural networks (ANNs), however, it continues to be a poorly understood phenomenon. Despite the extensive amount of work on catastrophic forgetting, we argue that it is still unclear how exactly the phenomenon should be quantified, and, moreover, to what degree all of the choices we make when designing learning systems affect the amount of catastrophic forgetting. We use various testbeds from the reinforcement learning and supervised learning literature to (1) provide evidence that the choice of which modern gradient-based optimization algorithm is used to train an ANN has a significant impact on the amount of catastrophic forgetting and show that-surprisingly-in many instances classical algorithms such as vanilla SGD experience less catastrophic forgetting than the more modern algorithms such as Adam. We empirically compare four different existing metrics for quantifying catastrophic forgetting and (2) show that the degree to which the learning systems experience catastrophic forgetting is sufficiently sensitive to the metric used that a change from one principled metric to another is enough to change the conclusions of a study dramatically. Our results suggest that a much more rigorous experimental methodology is required when looking at catastrophic forgetting. Based on our results, we recommend inter-task forgetting in supervised learning must be measured with both retention and relearning metrics concurrently, and intra-task forgetting in reinforcement learning must-at the very least-be measured with pairwise interference.

Chen Ma, Dylan R. Ashley, Junfeng Wen et al.

Transfer in Reinforcement Learning (RL) refers to the idea of applying knowledge gained from previous tasks to solving related tasks. Learning a universal value function (Schaul et al., 2015), which generalizes over goals and states, has previously been shown to be useful for transfer. However, successor features are believed to be more suitable than values for transfer (Dayan, 1993; Barreto et al.,2017), even though they cannot directly generalize to new goals. In this paper, we propose (1) Universal Successor Features (USFs) to capture the underlying dynamics of the environment while allowing generalization to unseen goals and (2) a flexible end-to-end model of USFs that can be trained by interacting with the environment. We show that learning USFs is compatible with any RL algorithm that learns state values using a temporal difference method. Our experiments in a simple gridworld and with two MuJoCo environments show that USFs can greatly accelerate training when learning multiple tasks and can effectively transfer knowledge to new tasks.

Craig Sherstan, Brendan Bennett, Kenny Young et al.

This paper investigates estimating the variance of a temporal-difference learning agent's update target. Most reinforcement learning methods use an estimate of the value function, which captures how good it is for the agent to be in a particular state and is mathematically expressed as the expected sum of discounted future rewards (called the return). These values can be straightforwardly estimated by averaging batches of returns using Monte Carlo methods. However, if we wish to update the agent's value estimates during learning--before terminal outcomes are observed--we must use a different estimation target called the λ-return, which truncates the return with the agent's own estimate of the value function. Temporal difference learning methods estimate the expected λ-return for each state, allowing these methods to update online and incrementally, and in most cases achieve better generalization error and faster learning than Monte Carlo methods. Naturally one could attempt to estimate higher-order moments of the λ-return. This paper is about estimating the variance of the λ-return. Prior work has shown that given estimates of the variance of the λ-return, learning systems can be constructed to (1) mitigate risk in action selection, and (2) automatically adapt the parameters of the learning process itself to improve performance. Unfortunately, existing methods for estimating the variance of the λ-return are complex and not well understood empirically. We contribute a method for estimating the variance of the λ-return directly using policy evaluation methods from reinforcement learning. Our approach is significantly simpler than prior methods that independently estimate the second moment of the λ-return. Empirically our new approach behaves at least as well as existing approaches, but is generally more robust.