Hyperparameter Optimisation with Practical Interpretability and Explanation Methods in Probabilistic Curriculum Learning

Hyperparameter optimisation (HPO) is crucial for achieving strong performance in reinforcement learning (RL), as RL algorithms are inherently sensitive to hyperparameter settings. Probabilistic Curriculum Learning (PCL) is a curriculum learning strategy designed to improve RL performance by structuring the agent's learning process, yet effective hyperparameter tuning remains challenging and computationally demanding. In this paper, we provide an empirical analysis of hyperparameter interactions and their effects on the performance of a PCL algorithm within standard RL tasks, including point-maze navigation and DC motor control. Using the AlgOS framework integrated with Optuna's Tree-Structured Parzen Estimator (TPE), we present strategies to refine hyperparameter search spaces, enhancing optimisation efficiency. Additionally, we introduce a novel SHAP-based interpretability approach tailored specifically for analysing hyperparameter impacts, offering clear insights into how individual hyperparameters and their interactions influence RL performance. Our work contributes practical guidelines and interpretability tools that significantly improve the effectiveness and computational feasibility of hyperparameter optimisation in reinforcement learning.

AlgOS: Algorithm Operating System

Algorithm Operating System (AlgOS) is an unopinionated, extensible, modular framework for algorithmic implementations. AlgOS offers numerous features: integration with Optuna for automated hyperparameter tuning; automated argument parsing for generic command-line interfaces; automated registration of new classes; and a centralised database for logging experiments and studies. These features are designed to reduce the overhead of implementing new algorithms and to standardise the comparison of algorithms. The standardisation of algorithmic implementations is crucial for reproducibility and reliability in research. AlgOS combines Abstract Syntax Trees with a novel implementation of the Observer pattern to control the logical flow of algorithmic segments.

Probabilistic Curriculum Learning for Goal-Based Reinforcement Learning

Reinforcement learning (RL) -- algorithms that teach artificial agents to interact with environments by maximising reward signals -- has achieved significant success in recent years. These successes have been facilitated by advances in algorithms (e.g., deep Q-learning, deep deterministic policy gradients, proximal policy optimisation, trust region policy optimisation, and soft actor-critic) and specialised computational resources such as GPUs and TPUs. One promising research direction involves introducing goals to allow multimodal policies, commonly through hierarchical or curriculum reinforcement learning. These methods systematically decompose complex behaviours into simpler sub-tasks, analogous to how humans progressively learn skills (e.g. we learn to run before we walk, or we learn arithmetic before calculus). However, fully automating goal creation remains an open challenge. We present a novel probabilistic curriculum learning algorithm to suggest goals for reinforcement learning agents in continuous control and navigation tasks.

Whole-Body Dynamic Throwing with Legged Manipulators

Most robotic behaviours focus on either manipulation or locomotion, where tasks that require the integration of both, such as full-body throwing, remain under-explored. Throwing with a robot involves complex coordination between object manipulation and legged locomotion, which is crucial for advanced real-world interactions. This work investigates the challenge of full-body throwing in robotic systems and highlights the advantages of utilising the robot's entire body. We propose a deep reinforcement learning (RL) approach that leverages the robot's body to enhance throwing performance through a strategically designed curriculum to avoid local optima and sparse but informative reward functions to improve policy flexibility. The robot's body learns to generate additional momentum and fine-tune the projectile release velocity. Our full-body method achieves on average 47% greater throwing distance and 34% greater throwing accuracy than the arm alone, across two robot morphologies - an armed quadruped and a humanoid. We also extend our method to optimise robot stability during throws. The learned policy effectively generalises throwing to targets at any 3D point in space within a specified range, which has not previously been achieved and does so with human-level throwing accuracy. We successfully transferred this approach from simulation to a real robot using sim2real techniques, demonstrating its practical viability.

Analyzing the Runtime of the Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA) on the Concatenated Trap Function

The Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA) is a state of the art evolutionary algorithm that leverages linkage learning to efficiently exploit problem structure. By identifying and preserving important building blocks during variation, GOMEA has shown promising performance on various optimization problems. In this paper, we provide the first runtime analysis of GOMEA on the concatenated trap function, a challenging benchmark problem that consists of multiple deceptive subfunctions. We derived an upper bound on the expected runtime of GOMEA with a truthful linkage model, showing that it can solve the problem in $O(m^{3}2^k)$ with high probability, where $m$ is the number of subfunctions and $k$ is the subfunction length. This is a significant speedup compared to the (1+1) EA, which requires $O(ln{(m)}(mk)^{k})$ expected evaluations.

Modularity based linkage model for neuroevolution

Crossover between neural networks is considered disruptive due to the strong functional dependency between connection weights. We propose a modularity-based linkage model at the weight level to preserve functionally dependent communities (building blocks) in neural networks during mixing. A proximity matrix is built by estimating the dependency between weights, then a community detection algorithm maximizing modularity is run on the graph described by such matrix. The resulting communities/groups of parameters are considered to be mutually independent and used as crossover masks in an optimal mixing EA. A variant is tested with an operator that neutralizes the permutation problem of neural networks to a degree. Experiments were performed on 8 and 10-bit parity problems as the intrinsic hierarchical nature of the dependencies in these problems are challenging to learn. The results show that our algorithm finds better, more functionally dependent linkage which leads to more successful crossover and better performance.

Pittsburgh Learning Classifier Systems for Explainable Reinforcement Learning: Comparing with XCS

Interest in reinforcement learning (RL) has recently surged due to the application of deep learning techniques, but these connectionist approaches are opaque compared with symbolic systems. Learning Classifier Systems (LCSs) are evolutionary machine learning systems that can be categorised as eXplainable AI (XAI) due to their rule-based nature. Michigan LCSs are commonly used in RL domains as the alternative Pittsburgh systems (e.g. SAMUEL) suffer from complex algorithmic design and high computational requirements; however they can produce more compact/interpretable solutions than Michigan systems. We aim to develop two novel Pittsburgh LCSs to address RL domains: PPL-DL and PPL-ST. The former acts as a "zeroth-level" system, and the latter revisits SAMUEL's core Monte Carlo learning mechanism for estimating rule strength. We compare our two Pittsburgh systems to the Michigan system XCS across deterministic and stochastic FrozenLake environments. Results show that PPL-ST performs on-par or better than PPL-DL and outperforms XCS in the presence of high levels of environmental uncertainty. Rulesets evolved by PPL-ST can achieve higher performance than those evolved by XCS, but in a more parsimonious and therefore more interpretable fashion, albeit with higher computational cost. This indicates that PPL-ST is an LCS well-suited to producing explainable policies in RL domains.

A Genetic Fuzzy System for Interpretable and Parsimonious Reinforcement Learning Policies

Reinforcement learning (RL) is experiencing a resurgence in research interest, where Learning Classifier Systems (LCSs) have been applied for many years. However, traditional Michigan approaches tend to evolve large rule bases that are difficult to interpret or scale to domains beyond standard mazes. A Pittsburgh Genetic Fuzzy System (dubbed Fuzzy MoCoCo) is proposed that utilises both multiobjective and cooperative coevolutionary mechanisms to evolve fuzzy rule-based policies for RL environments. Multiobjectivity in the system is concerned with policy performance vs. complexity. The continuous state RL environment Mountain Car is used as a testing bed for the proposed system. Results show the system is able to effectively explore the trade-off between policy performance and complexity, and learn interpretable, high-performing policies that use as few rules as possible.

Modularity in NEAT Reinforcement Learning Networks

Modularity is essential to many well-performing structured systems, as it is a useful means of managing complexity [8]. An analysis of modularity in neural networks produced by machine learning algorithms can offer valuable insight into the workings of such algorithms and how modularity can be leveraged to improve performance. However, this property is often overlooked in the neuroevolutionary literature, so the modular nature of many learning algorithms is unknown. This property was assessed on the popular algorithm "NeuroEvolution of Augmenting Topologies" (NEAT) for standard simulation benchmark control problems due to NEAT's ability to optimise network topology. This paper shows that NEAT networks seem to rapidly increase in modularity over time with the rate and convergence dependent on the problem. Interestingly, NEAT tends towards increasingly modular networks even when network fitness converges. It was shown that the ideal level of network modularity in the explored parameter space is highly dependent on other network variables, dispelling theories that modularity has a straightforward relationship to network performance. This is further proven in this paper by demonstrating that rewarding modularity directly did not improve fitness.

Approximate discounting-free policy evaluation from transient and recurrent states

In order to distinguish policies that prescribe good from bad actions in transient states, we need to evaluate the so-called bias of a policy from transient states. However, we observe that most (if not all) works in approximate discounting-free policy evaluation thus far are developed for estimating the bias solely from recurrent states. We therefore propose a system of approximators for the bias (specifically, its relative value) from transient and recurrent states. Its key ingredient is a seminorm LSTD (least-squares temporal difference), for which we derive its minimizer expression that enables approximation by sampling required in model-free reinforcement learning. This seminorm LSTD also facilitates the formulation of a general unifying procedure for LSTD-based policy value approximators. Experimental results validate the effectiveness of our proposed method.