Deep Learning for Reinforcement Learning
University of St Andrews
Abstract
Dynamic Algorithm Configuration (DAC) studies the efficient identification of control policies for parameterized optimization algorithms. Numerous studies have leveraged the robustness of decision-making in Reinforcement Learning (RL) to address the optimization challenges in algorithm configuration. However, applying RL to DAC is challenging and often requires extensive domain expertise. We conduct a comprehensive study of deep-RL algorithms in DAC through a systematic analysis of controlling the population size parameter of the (1+($λ$,$λ$))-GA on OneMax instances. Our investigation of DDQN and PPO reveals two fundamental challenges that limit their effectiveness in DAC: scalability degradation and learning instability. We trace these issues to two primary causes: under-exploration and planning horizon coverage, each of which can be effectively addressed through targeted solutions. To address under-exploration, we introduce an adaptive reward shifting mechanism that leverages reward distribution statistics to enhance DDQN agent exploration, eliminating the need for instance-specific hyperparameter tuning and ensuring consistent effectiveness across different problem scales. In dealing with the planning horizon coverage problem, we demonstrate that undiscounted learning effectively resolves it in DDQN, while PPO faces fundamental variance issues that necessitate alternative algorithmic designs. We further analyze the hyperparameter dependencies of PPO, showing that while hyperparameter optimization enhances learning stability, it consistently falls short in identifying effective policies across various configurations. Finally, we demonstrate that DDQN equipped with our adaptive reward shifting strategy achieves performance comparable to theoretically derived policies with vastly improved sample efficiency, outperforming prior DAC approaches by several orders of magnitude.
Dynamic Algorithm Configuration (DAC) studies the efficient identification of control policies for parameterized optimization algorithms. Numerous studies have leveraged the robustness of decision-making in Reinforcement Learning (RL) to address the optimization challenges in algorithm configuration. However, applying RL to DAC is challenging and often requires extensive domain expertise. We conduct a comprehensive study of deep-RL algorithms in DAC through a systematic analysis of controlling the population size parameter of the (1+($λ$,$λ$))-GA on OneMax instances. Our investigation of DDQN and PPO reveals two fundamental challenges that limit their effectiveness in DAC: scalability degradation and learning instability. We trace these issues to two primary causes: under-exploration and planning horizon coverage, each of which can be effectively addressed through targeted solutions. To address under-exploration, we introduce an adaptive reward shifting mechanism that leverages reward distribution statistics to enhance DDQN agent exploration, eliminating the need for instance-specific hyperparameter tuning and ensuring consistent effectiveness across different problem scales. In dealing with the planning horizon coverage problem, we demonstrate that undiscounted learning effectively resolves it in DDQN, while PPO faces fundamental variance issues that necessitate alternative algorithmic designs. We further analyze the hyperparameter dependencies of PPO, showing that while hyperparameter optimization enhances learning stability, it consistently falls short in identifying effective policies across various configurations. Finally, we demonstrate that DDQN equipped with our adaptive reward shifting strategy achieves performance comparable to theoretically derived policies with vastly improved sample efficiency, outperforming prior DAC approaches by several orders of magnitude.
AI Summary
- Entropy regularization is used to encourage exploration, but it fails to provide sufficient guidance for discovering and stabilizing on better policies. [3]
- The paper suggests that relying solely on the ε-greedy strategy is inadequate to address under-exploration. [3]
- The effectiveness of reward shaping lies in its ability to accelerate training by minimizing unnecessary steps within an episode. [3]
- Entropy regularization: a technique used to encourage exploration by adding an intrinsic reward term to the original objective. [3]
- Reward shaping: a strategy that incorporates an external reward factor into the original reward to stabilize and accelerate the training process. [3]
- PPO (Proximal Policy Optimization): a model-free reinforcement learning algorithm that uses trust region optimization to update the policy. [3]
- ε-greedy: a simple exploration strategy where the agent chooses between exploiting the current state or exploring a new one with probability ε. [3]
- The paper discusses the challenges faced by PPO in a dynamic algorithm configuration environment, specifically under-exploration. [2]
Sapienza University of
Abstract
The main focus of Hierarchical Reinforcement Learning (HRL) is studying how large Markov Decision Processes (MDPs) can be more efficiently solved when addressed in a modular way, by combining partial solutions computed for smaller subtasks. Despite their very intuitive role for learning, most notions of MDP abstractions proposed in the HRL literature have limited expressive power or do not possess formal efficiency guarantees. This work addresses these fundamental issues by defining Realizable Abstractions, a new relation between generic low-level MDPs and their associated high-level decision processes. The notion we propose avoids non-Markovianity issues and has desirable near-optimality guarantees. Indeed, we show that any abstract policy for Realizable Abstractions can be translated into near-optimal policies for the low-level MDP, through a suitable composition of options. As demonstrated in the paper, these options can be expressed as solutions of specific constrained MDPs. Based on these findings, we propose RARL, a new HRL algorithm that returns compositional and near-optimal low-level policies, taking advantage of the Realizable Abstraction given in the input. We show that RARL is Probably Approximately Correct, it converges in a polynomial number of samples, and it is robust to inaccuracies in the abstraction.
The main focus of Hierarchical Reinforcement Learning (HRL) is studying how large Markov Decision Processes (MDPs) can be more efficiently solved when addressed in a modular way, by combining partial solutions computed for smaller subtasks. Despite their very intuitive role for learning, most notions of MDP abstractions proposed in the HRL literature have limited expressive power or do not possess formal efficiency guarantees. This work addresses these fundamental issues by defining Realizable Abstractions, a new relation between generic low-level MDPs and their associated high-level decision processes. The notion we propose avoids non-Markovianity issues and has desirable near-optimality guarantees. Indeed, we show that any abstract policy for Realizable Abstractions can be translated into near-optimal policies for the low-level MDP, through a suitable composition of options. As demonstrated in the paper, these options can be expressed as solutions of specific constrained MDPs. Based on these findings, we propose RARL, a new HRL algorithm that returns compositional and near-optimal low-level policies, taking advantage of the Realizable Abstraction given in the input. We show that RARL is Probably Approximately Correct, it converges in a polynomial number of samples, and it is robust to inaccuracies in the abstraction.
Reinforcement Learning
ANU
Abstract
This paper discusses theoretical and experimental aspects of gradient-based approaches to the direct optimization of policy performance in controlled POMDPs. We introduce GPOMDP, a REINFORCE-like algorithm for estimating an approximation to the gradient of the average reward as a function of the parameters of a stochastic policy. The algorithm's chief advantages are that it requires only a single sample path of the underlying Markov chain, it uses only one free parameter $β\in [0,1)$, which has a natural interpretation in terms of bias-variance trade-off, and it requires no knowledge of the underlying state. We prove convergence of GPOMDP and show how the gradient estimates produced by GPOMDP can be used in a conjugate-gradient procedure to find local optima of the average reward.
This paper discusses theoretical and experimental aspects of gradient-based approaches to the direct optimization of policy performance in controlled POMDPs. We introduce GPOMDP, a REINFORCE-like algorithm for estimating an approximation to the gradient of the average reward as a function of the parameters of a stochastic policy. The algorithm's chief advantages are that it requires only a single sample path of the underlying Markov chain, it uses only one free parameter $β\in [0,1)$, which has a natural interpretation in terms of bias-variance trade-off, and it requires no knowledge of the underlying state. We prove convergence of GPOMDP and show how the gradient estimates produced by GPOMDP can be used in a conjugate-gradient procedure to find local optima of the average reward.
AI Summary
- The authors analyze an algorithm for computing an approximation to the performance gradient and demonstrate its use in a greedy local search algorithm. [3]
- They also discuss related work and future research directions. [3]
- The need to specify running times and the discount factor β in advance Gradient-based methods have been shown to be effective for solving POMDPs, but they can suffer from high variance. [2]
University of Waterloo
Abstract
This paper critiques common patterns in machine ethics for Reinforcement Learning (RL) and argues for a virtue focused alternative. We highlight two recurring limitations in much of the current literature: (i) rule based (deontological) methods that encode duties as constraints or shields often struggle under ambiguity and nonstationarity and do not cultivate lasting habits, and (ii) many reward based approaches, especially single objective RL, implicitly compress diverse moral considerations into a single scalar signal, which can obscure trade offs and invite proxy gaming in practice. We instead treat ethics as policy level dispositions, that is, relatively stable habits that hold up when incentives, partners, or contexts change. This shifts evaluation beyond rule checks or scalar returns toward trait summaries, durability under interventions, and explicit reporting of moral trade offs. Our roadmap combines four components: (1) social learning in multi agent RL to acquire virtue like patterns from imperfect but normatively informed exemplars; (2) multi objective and constrained formulations that preserve value conflicts and incorporate risk aware criteria to guard against harm; (3) affinity based regularization toward updateable virtue priors that support trait like stability under distribution shift while allowing norms to evolve; and (4) operationalizing diverse ethical traditions as practical control signals, making explicit the value and cultural assumptions that shape ethical RL benchmarks.
This paper critiques common patterns in machine ethics for Reinforcement Learning (RL) and argues for a virtue focused alternative. We highlight two recurring limitations in much of the current literature: (i) rule based (deontological) methods that encode duties as constraints or shields often struggle under ambiguity and nonstationarity and do not cultivate lasting habits, and (ii) many reward based approaches, especially single objective RL, implicitly compress diverse moral considerations into a single scalar signal, which can obscure trade offs and invite proxy gaming in practice. We instead treat ethics as policy level dispositions, that is, relatively stable habits that hold up when incentives, partners, or contexts change. This shifts evaluation beyond rule checks or scalar returns toward trait summaries, durability under interventions, and explicit reporting of moral trade offs. Our roadmap combines four components: (1) social learning in multi agent RL to acquire virtue like patterns from imperfect but normatively informed exemplars; (2) multi objective and constrained formulations that preserve value conflicts and incorporate risk aware criteria to guard against harm; (3) affinity based regularization toward updateable virtue priors that support trait like stability under distribution shift while allowing norms to evolve; and (4) operationalizing diverse ethical traditions as practical control signals, making explicit the value and cultural assumptions that shape ethical RL benchmarks.
Agentic RL
University of Illinois at
Abstract
Agentic systems are modern software systems: they consist of orchestrated modules, expose interfaces, and are deployed in software pipelines. Unlike conventional programs, their execution (i.e., trajectories) is inherently stochastic and adaptive to the problem they are solving. Evaluation of such systems is often outcome-centric, judging their performance based on success or failure at the final step. This narrow focus overlooks detailed insights about such systems, failing to explain how agents reason, plan, act, or change their strategies over time. Inspired by the structured representation of conventional software systems as graphs, we introduce Graphectory to systematically encode the temporal and semantic relations in such software systems. Graphectory facilitates the design of process-centric metrics and analyses to assess the quality of agentic workflows independent of final success. Using Graphectory, we analyze 4000 trajectories of two dominant agentic programming workflows, namely SWE-agent and OpenHands, with a combination of four backbone Large Language Models (LLMs), attempting to resolve SWE-bench Verified issues. Our fully automated analyses reveal that: (1) agents using richer prompts or stronger LLMs exhibit more complex Graphectory, reflecting deeper exploration, broader context gathering, and more thorough validation before patch submission; (2) agents' problem-solving strategies vary with both problem difficulty and the underlying LLM -- for resolved issues, the strategies often follow coherent localization-patching-validation steps, while unresolved ones exhibit chaotic, repetitive, or backtracking behaviors; (3) even when successful, agentic programming systems often display inefficient processes, leading to unnecessarily prolonged trajectories.
Agentic systems are modern software systems: they consist of orchestrated modules, expose interfaces, and are deployed in software pipelines. Unlike conventional programs, their execution (i.e., trajectories) is inherently stochastic and adaptive to the problem they are solving. Evaluation of such systems is often outcome-centric, judging their performance based on success or failure at the final step. This narrow focus overlooks detailed insights about such systems, failing to explain how agents reason, plan, act, or change their strategies over time. Inspired by the structured representation of conventional software systems as graphs, we introduce Graphectory to systematically encode the temporal and semantic relations in such software systems. Graphectory facilitates the design of process-centric metrics and analyses to assess the quality of agentic workflows independent of final success. Using Graphectory, we analyze 4000 trajectories of two dominant agentic programming workflows, namely SWE-agent and OpenHands, with a combination of four backbone Large Language Models (LLMs), attempting to resolve SWE-bench Verified issues. Our fully automated analyses reveal that: (1) agents using richer prompts or stronger LLMs exhibit more complex Graphectory, reflecting deeper exploration, broader context gathering, and more thorough validation before patch submission; (2) agents' problem-solving strategies vary with both problem difficulty and the underlying LLM -- for resolved issues, the strategies often follow coherent localization-patching-validation steps, while unresolved ones exhibit chaotic, repetitive, or backtracking behaviors; (3) even when successful, agentic programming systems often display inefficient processes, leading to unnecessarily prolonged trajectories.
AI Summary
- The study also finds a correlation between process-centric metrics and repair status, indicating that agents with higher values in certain metrics are more likely to resolve issues. [3]
- The research questions focus on process-centric metrics, problem-solving strategies, and inefficiency patterns in programming agents. [2]
- The study explores the use of graph-based structures, Graphectory and Langutury, to analyze the behavior of programming agents. [1]
IBM
Abstract
The rapid deployment of large language model (LLM)-based agents introduces a new class of risks, driven by their capacity for autonomous planning, multi-step tool integration, and emergent interactions. It raises some risk factors for existing governance approaches as they remain fragmented: Existing frameworks are either static taxonomies driven; however, they lack an integrated end-to-end pipeline from risk identification to operational assurance, especially for an agentic platform. We propose AGENTSAFE, a practical governance framework for LLM-based agentic systems. The framework operationalises the AI Risk Repository into design, runtime, and audit controls, offering a governance framework for risk identification and assurance. The proposed framework, AGENTSAFE, profiles agentic loops (plan -> act -> observe -> reflect) and toolchains, and maps risks onto structured taxonomies extended with agent-specific vulnerabilities. It introduces safeguards that constrain risky behaviours, escalates high-impact actions to human oversight, and evaluates systems through pre-deployment scenario banks spanning security, privacy, fairness, and systemic safety. During deployment, AGENTSAFE ensures continuous governance through semantic telemetry, dynamic authorization, anomaly detection, and interruptibility mechanisms. Provenance and accountability are reinforced through cryptographic tracing and organizational controls, enabling measurable, auditable assurance across the lifecycle of agentic AI systems. The key contributions of this paper are: (1) a unified governance framework that translates risk taxonomies into actionable design, runtime, and audit controls; (2) an Agent Safety Evaluation methodology that provides measurable pre-deployment assurance; and (3) a set of runtime governance and accountability mechanisms that institutionalise trust in agentic AI ecosystems.
The rapid deployment of large language model (LLM)-based agents introduces a new class of risks, driven by their capacity for autonomous planning, multi-step tool integration, and emergent interactions. It raises some risk factors for existing governance approaches as they remain fragmented: Existing frameworks are either static taxonomies driven; however, they lack an integrated end-to-end pipeline from risk identification to operational assurance, especially for an agentic platform. We propose AGENTSAFE, a practical governance framework for LLM-based agentic systems. The framework operationalises the AI Risk Repository into design, runtime, and audit controls, offering a governance framework for risk identification and assurance. The proposed framework, AGENTSAFE, profiles agentic loops (plan -> act -> observe -> reflect) and toolchains, and maps risks onto structured taxonomies extended with agent-specific vulnerabilities. It introduces safeguards that constrain risky behaviours, escalates high-impact actions to human oversight, and evaluates systems through pre-deployment scenario banks spanning security, privacy, fairness, and systemic safety. During deployment, AGENTSAFE ensures continuous governance through semantic telemetry, dynamic authorization, anomaly detection, and interruptibility mechanisms. Provenance and accountability are reinforced through cryptographic tracing and organizational controls, enabling measurable, auditable assurance across the lifecycle of agentic AI systems. The key contributions of this paper are: (1) a unified governance framework that translates risk taxonomies into actionable design, runtime, and audit controls; (2) an Agent Safety Evaluation methodology that provides measurable pre-deployment assurance; and (3) a set of runtime governance and accountability mechanisms that institutionalise trust in agentic AI ecosystems.
AI Summary
- AGENTSAFE is an ethics-grounded governance framework that translates abstract safety principles into concrete, testable, and auditable practices. [2]
Help us improve your experience!
This project is on its early stages your feedback can be pivotal on the future of the project.
Let us know what you think about this week's papers and suggestions!
Give Feedback