Global ETD Search
Search theses and dissertations gathered from participating repositories worldwide. Every result links back to the library that holds it. No account is needed.
Results
Showing 1 to 20 of 37 for “"Reward Models"”.
-
Foundation Reward Models for Robot Learning
… in imitation learning or carefully engineered reward functions in reinforcement learning. This dependency limits scalability, as it is infeasible to collect demonstrations or design reward functions for every possible task and environment. This thesis proposes a viable path towards scaling …
-
Construction and solution of Markov reward models
… can be defined at the net level by means of reward variables. Depending on the measures, these net-level reward models are solved either by first generating a state-level reward model or by directly generating paths from the net-level description. In this thesis, we propose algorithms for the …
-
Algorithms for reachability problems on stochastic Markov reward models
… to automate the formal analysis of probabilistic models such as Markov chains. In this thesis, we study and develop the stochastic Markov reward model (sMRM) which extends the Markov chain with rewards as random variables. The model recently being introduced, does not have much in the way of …
-
Generating humor with information-theoretic reward signals - A system trained to write funny cartoon captions and humorous news headlines, optimized with reinforcement learning using human preference reward models and an information-theoretic approach to humor theory (:
… yet it remains difficult for large language models to both generate and evaluate. This thesis investigates whether information-theoretic reward signals, paired with a trained reward model, can align a language model to generate funny outputs. We frame humor generation as a reinforcement …
-
Renewable electric power systems energy yield and performance estimation
This dissertation presents models for reliability assessment, energy yield estimation, and uncertainty analysis of renewable electric power systems. We propose system performability models that describe system attributes while acknowledging failures and repairs in constituent elements. Two broad …
-
Designing, learning, and utilizing dense reward for generative AI alignment
… the gap between concrete task domain and reward signal for policy training, with a particular focus on how to design, learn, and utilize dense reward for reinforcement-learning-from-human-feedback (RLHF) training of text and text-to-image generation models. We start from learning better …
-
Exploring Optimization Strategies for Multi-Objective Reward Modeling
To align Large language models with human preferences effectively, a framework called Reinforcement Learning from Human Feedback has become a standard for this purpose. In the RLHF, reward models play an impotent role. It was used to approximate human judgment, the reward model will provide a …
-
Toward Reliable Supervision for Foundation Models
Foundation models have made rapid progress in reasoning, alignment, and multimodal understanding, but their reliability critically depends on the quality of supervision used during learning. Such supervision may come from learned reward models, verifiable outcomes, expert models, the model itself, …
-
Instruction Mining from Images: Constructing a Synthetic Dataset for Multimodal Learning
… των μεγάλων γλωσσικών μοντέλων (Large Language Models – LLMs) και των πολυτροπικών μοντέλων όρασης–γλώσσας (Multimodal Large Language Models – MLLMs) έχει οδηγήσει στην ανάπτυξη συστημάτων ικανών να συνδυάζουν οπτική και γλωσσική πληροφορία για την εκτέλεση σύνθετων εργασιών συλλογισμού, …
-
VISION-BASED DEEP REINFORCEMENT LEARNING FOR ROBOTIC MANIPULATION
… A Bradley–Terry gradient analysis shows that reward models learn only from comparisons that remain uncertain and that the labeler can reliably judge. On Lift, an expert with 20% success outperforms all stronger experts, reaching 76.2% success under VLM supervision, whereas on the …
-
Robust and Flexible Reward Modeling for LLM Alignment
As large language models grow increasingly more capable, ensuring their alignment with human values is of utmost importance. One of the most promising ways to align language models is by designing a reward function that measures alignment with human values, and training the language model to …
-
COMPOSITIONAL REASONING AND PLANNING WITH FOUNDATION MODELS
Foundation models are pre-trained by internet-scale data and possess vast knowledge. However, they struggle with compositional tasks in reasoning and planning. In those tasks, evidence is scattered across multiple sources and needs to be considered together simultaneously or connected in a …
-
MULTI-ROBOT COORDINATION AND COOPERATION VIA GRAPH-BASED COMPUTATION
… in representing inter-region relationships and reward models. We first address the multi-agent coverage control problem, introducing graph-based computation via Graph Neural Networks, which boasts improved performance and scalability by leveraging learned inter-agent communication strategies. We …
-
Advancing reasoning and planning in large language models via reward shaping
Recent advancements in large language models (LLMs) have significantly enhanced their reasoning and planning capabilities, enabling them to serve effectively in complex, real-world scenarios. Despite these improvements, achieving human-level performance remains challenging, particularly for tasks …
-
Theory and Practice of Large-scale Logistics: Offline Contextual Bandits and Decomposition Methods
… settings. Unlike standard methods that fail when reward models are poorly specified, ESR provably yields policies that asymptotically achieve optimal performance while remaining compatible with gradient-based optimization. The value of this approach is demonstrated through applications in health …
-
Optimization of protocols
… are specified informally and implicitly by reward models trained on data. A mathematical model of protocol post-training is proposed in terms of an objective function defined by such rewards and regularized by statistical distances from the pre-trained behaviors. It is shown that any …
-
Scaling Human Supervision for Robotic Manipulation
… methods for distilling human supervision into reward models that enable reinforcement learning beyond the original data. These learned rewards allow robots to refine their behavior autonomously, increasing sample efficiency while reducing dependence on constant human input. Second, it explores …
-
Safe and transferable multi-task bandit learning with shared representations and safe reinforcement learning
… the low-rank structure of the underlying reward models to design fast, sample-efficient algorithms with provable guarantees. In the second part of this thesis, we study transfer across tasks by meta-learning and task relevance, specifically by learning a shared low-rank representation from …
-
Unmanned Aerial Vehicle Autonomous Navigation using Semi-Prior Deep Reinforcement Learning
… address the above issues, I propose an cumulated reward model in this thesis to overcome the sparsity of the training data for DRL algorithms, and propose a region segmentation algorithm to reduce the probability of DRL algorithms UAV agents falling into the local optimal trap. Furthermore, I …
-
Deliberative Machines: From Reflective Dialogue to Fair Consensus with Language Models and Social Choice
… intelligence (AI), particularly large language models (LLMs), and social choice theory. Firstly, it explores how principles from social choice can address challenges in AI alignment, specifically the problem of aggregating diverse human preferences fairly when guiding AI behavior (SC → AI). …
Page 1 of 2