School of Computer Science and Applied Mathematics (ETDs)

Permanent URI for this communityhttps://hdl.handle.net/10539/38004

Browse

Has files: YesSubject: search.filters.subject.Reinforcement Learning

Search Results

Now showing 1 - 5 of 5
  • No Thumbnail Available
    Item
    Towards Lifelong Reinforcement Learning through Temporal Logics and Zero-Shot Composition
    (2024-10) Tasse, Geraud Nangue; Rosman, Benjamin; James, Steven
    This thesis addresses the fundamental challenge of creating agents capable of solving a wide range of tasks in their environments, akin to human capabilities. For such agents to be truly useful and be capable of assisting humans in our day-to-day lives, we identify three key abilities that general purpose agents should have: Flexibility, Instructability, and Reliability (FIRe). Flexibility refers to the ability of agents to adapt to various tasks with minimal learning; instructability involves the capacity for agents to understand and execute task specifications provided by humans in a comprehensible manner; and reliability entails agents’ ability to solve tasks safely and effectively with theoretical guarantees on their behavior. To build such agents, reinforcement learning (RL) is the framework of choice given that it is the only one that models the agent-environment interaction. It is also particularly promising since it has shown remarkable success in recent years in various domains—including gaming, scientific research, and robotic control. However, prevailing RL methods often fall short of the FIRe desiderata. They typically exhibit poor sample efficiency, demanding millions of environment interactions to learn optimal behaviors. Task specification relies heavily on hand-designed reward functions, posing challenges for non-experts in defining tasks. Moreover, these methods tend to specialize in single tasks, lacking guarantees on the broader adaptability and behavior robustness desired for lifelong agents that need solve multiple tasks. Clearly, the regular RL framework is not enough, and does not capture important aspects of what makes humans so general—such as the use of language to specify and understand tasks. To address these shortcomings, we propose a principled framework for the logical composition of arbitrary tasks in an environment, and introduce a novel knowledge representation called World Value Functions (WVFs) that will enable agents to solve arbitrary tasks specified using language. The use of logical composition is inspired by the fact that all formal languages are built upon the rules of propositional logics. Hence, if we want agents that understand tasks specified in any formal language, we must define what it means to apply the usual logic operators (conjunction, disjunction, and negation) over tasks. The introduction of WVFs is inspired by the fact that humans seem to always seek general knowledge about how to achieve a variety of goals in their environment, irrespective of the specific task they are learning. Our main contributions include: (i) Instructable agents: We formalize the logical composition of arbitrary tasks in potentially stochastic environments, and ensure that task compositions lead to rewards minimising undesired behaviors. (ii) Flexible agents: We introduce WVFs as a new objective for RL agents, enabling them to solve a variety of tasks in their environment. Additionally, we demonstrate zero-shot skill composition and lifelong sample efficiency. (iii) Reliable agents: We develop methods for agents to understand and execute both natural and formal language instructions, ensuring correctness and safety in task execution, particularly in real-world scenarios. By addressing these challenges, our framework represents a significant step towards achieving the FIRe desiderata in AI agents, thereby enhancing their utility and safety in a lifelong learning setting like the real world.
  • Thumbnail Image
    Item
    Counting Reward Automata: Exploiting Structure in Reward Functions Expressible in Decidable Formal Languages
    (University of the Witwatersrand, Johannesburg, 2024-07) Bester, Tristan; Rosman, Benjamin; James, Steven; Tasse, Geraud Nangue
    In general, reinforcement learning agents are restricted from directly accessing the environment model. This restricts the agent’s access to the environmental dynamics and reward models, which are only accessible through repeated environmental interactions. As reinforcement learning is well suited for use in complex environments, which are challenging to model, the general assumption that the transition probabilities associated with the environment are unknown is justified. However, as agents cannot discern rewards directly from the environment, reward functions must be designed and implemented for both simulated and real-world environments. As a result, the assumption that the reward model must remain hidden from the agent is unnecessary and detrimental to learning. Previously, methods have been developed that utilise the structure of the reward function to enable more sample-efficient learning. These methods employ a finite state machine variant to facilitate reward specification in a manner that exposes the internal structure of the reward function. This approach is particularly effective when solving long-horizon tasks as it enables the use of counterfactual reasoning with off-policy learning which significantly improves sample efficiency. However, as these approaches are dependent on finite-state machines, they are only able to express a small number of reward functions. This severely limits the applicability of these approaches as they cannot model simple tasks such as “fetch a coffee for each person in the office” which involves counting – one of the numerous properties finite state machines cannot model. This work addresses the limited expressiveness of current state machine-based approaches to reward modelling. Specifically, we introduce a novel approach compatible with any reward function which can be expressed as a well-defined algorithm We present the counting reward automaton – an abstract machine capable of modelling reward functions expressible in any decidable formal language. Unlike previous approaches to state machine-based reward modelling, which are limited to the expression of tasks as regular languages, our framework allows for tasks described by decidable formal languages. It follows that our framework is an extremely general approach to reward modelling – compatible with any task specification expressible as a well-defined algorithm. This is a significant contribution as it greatly extends the class of problems which can benefit from the improved learning techniques facilitated by state machine-based reward modelling. We prove that an agent equipped with such an abstract machine is able to solve an extended set of tasks. We show that this increase in expressive power does not come at the cost of increased automaton complexity. This is followed by the introduction of several learning algorithms designed to increase sample efficiency through the exploitation of automaton structure. These algorithms are based on counterfactual reasoning with off-policy RL and use techniques from the fields of HRL and reward shaping. Finally, we evaluate our approach in several domains requiring long-horizon plans. Empirical results demonstrate that our method outperforms competing approaches in terms of automaton complexity, sample efficiency, and task completion.
  • Thumbnail Image
    Item
    Creating an adaptive collaborative playstyle-aware companion agent
    (University of the Witwatersrand, Johannesburg, 2023-09) Arendse, Lindsay John; Rosman, Benjamin
    Companion characters in video games play a unique part in enriching player experience. Companion agents support the player as an ally or sidekick and would typically help the player by providing hints, resources, or even fight along-side the human player. Players often adopt a certain approach or strategy, referred to as a playstyle, whilst playing video games. Players do not only approach challenges in games differently, but also play games differently based on what they find rewarding. Companion agent characters thus have an important role to play by assisting the player in a way which aligns with their playstyle. Existing companion agent approaches fall short and adversely affect the collaborative experience when the companion agent is not able to assist the human player in a manner consistent with their playstyle. Furthermore, if the companion agent cannot assist in real time, player engagement levels are lowered since the player will need to wait for the agent to compute its action - leading to a frustrating player experience. We therefore present a framework for creating companion agents that are adaptive such that they respond in real time with actions that align with the player’s playstyle. Companion agents able to do so are what we refer to as playstyle-aware. Creating a playstyle-aware adaptive agent firstly requires a mechanism for correctly classifying or identifying the player style, before attempting to assist the player with a given task. We present a method which can enable the real time in-game playstyle classification of players. We contribute a hybrid probabilistic supervised learning framework, using Bayesian Inference informed by a K-Nearest Neighbours based likelihood, that is able to classify players in real time at every step within a given game level using only the latest player action or state observation. We empirically evaluate our hybrid classifier against existing work using MiniDungeons, a common benchmark game domain. We further evaluate our approach using real player data from the game Super Mario Bros. We out perform our comparative study and our results highlight the success of our framework in identifying playstyles in a complex human player setting. The second problem we explore is the problem of assisting the identified playstyle with a suitable action. We formally define this as the ‘Learning to Assist’ problem, where given a set of companion agent policies, we aim to determine the policy which best complements the observed playstyle. An action is complementary such that it aligns with the goal of the playstyle. We extend MiniDungeons into a two-player game called Collaborative MiniDungeons which we use to evaluate our companion agent against several comparative baselines. The results from this experiment highlights that companion agents which are able to adapt and assist different playstyles on average bring about a greater player experience when using a playstyle specific reward function as a proxy for what the players find rewarding. In this way we present an approach for creating adaptive companion agents which are playstyle-aware and able to collaborate with players in real time.
  • Thumbnail Image
    Item
    A Continuous Reinforcement Learning Approach to Self-Adaptive Particle Swarm Optimisation
    (University of the Witwatersrand, Johannesburg, 2023-08) Tilley, Duncan; Cleghorn, Christopher
    Particle Swarm Optimisation (PSO) is a popular black-box optimisation technique due to its simple implementation and surprising ability to perform well on various problems. Unfortunately, PSO is fairly sensitive to the choice of hyper-parameters. For this reason, many self-adaptive techniques have been proposed that attempt to both simplify hyper-parameter selection and improve the performance of PSO. Surveys however show that many self-adaptive techniques are still outperformed by time-varying techniques where the value of coefficients are simply increased or decreased over time. More recent works have shown the successful application of Reinforcement Learning (RL) to learn self-adaptive control policies for optimisers such as differential evolution, genetic algorithms, and PSO. However, many of these applications were limited to only discrete state and action spaces, which severely limits the choices available to a control policy, given that the PSO coefficients are continuous variables. This dissertation therefore investigates the application of continuous RL techniques to learn a self-adaptive control policy that can make full use of the continuous nature of the PSO coefficients. The dissertation first introduces the RL framework used to learn a continuous control policy by defining the environment, action-space, state-space, and a number of possible reward functions. An effective learning environment that is able to overcome the difficulties of continuous RL is then derived through a series of experiments, culminating in a successfully learned continuous control policy. The policy is then shown to perform well on the benchmark problems used during training when compared to other self-adaptive PSO algorithms. Further testing on benchmark problems not seen during training suggest that the learned policy may however not generalise well to other functions, but this is shown to also be a problem in other PSO algorithms. Finally, the dissertation performs a number of experiments to provide insights into the behaviours learned by the continuous control policy.
  • Thumbnail Image
    Item
    A fully-decentralised general-sum approach for multi-agent reinforcement learning using minimal modelling
    (University of the Witwatersrand, Johannesburg, 2023-08) Kruger, Marcel Matthew Anthony; Rosman, Benjamin; James, Steven; Shipton, Jarrod
    Multi-agent reinforcement learning is a prominent area of research in machine learning, extending reinforcement learning to scenarios where multiple agents concurrently learn and interact within the same environment. Most existing methods rely on centralisation during training, while others employ agent modelling. In contrast, we propose a novel method that adapts the role of entropy to assist in fully-decentralised training without explicitly modelling other agents using additional information to which most centralised methods assume access. We augment entropy to encourage more deterministic agents, and instead, we let the non-stationarity inherent in MARL serve as a mode for exploration. We empirically evaluate the performance of our method across five distinct environments, each representing unique challenges. Our assessment encompasses both cooperative and competitive cases. Our findings indicate that the approach of penalising entropy, rather than rewarding it, enables agents to perform at least as well as the prevailing standard of entropy maximisation. Moreover, our alternative approach achieves several of the original objectives of entropy regularisation in reinforcement learning, such as increased sample efficiency and potentially better final rewards. Whilst entropy has a significant role, our results in the competitive case indicate that position bias is still a considerable challenge.