Ai Reinforcement Learning

Reinforcement Learning Foundations

Doctoral work examines agents learning behaviour through interaction and feedback. Learning from consequences suits problems where correct answers are unknown.

Markov Decision Process Research

Research examines the mathematical framework underlying sequential decision problems. This framework is the formal foundation of nearly all work here.

Partially Observable Setting

Doctoral study examines decisions where the true situation is never fully visible. Partial observation describes nearly every realistic application.

State Representation Research

Research examines how situations should be described for effective learning. Representation quality strongly determines what an agent can learn.

Action Space Research

Doctoral work examines the structure of choices available to a learning agent. Action structure determines which algorithms are even applicable.

Continuous Control Research

Research examines problems where actions take smoothly varying values. Continuous settings dominate robotics and physical control applications.

Discrete Action Research

Doctoral study examines problems with finitely many distinct available choices. Discrete settings permit methods that continuous problems cannot use.

Reward Function Research

Research examines the signal defining what an agent should actually pursue. The reward function encodes the entire objective being optimised.

Reward Design Research

Doctoral work examines specifying objectives that produce the intended behaviour. Poorly specified objectives are among the commonest failure sources.

Reward Shaping Research

Research examines supplementary signals guiding learning toward useful behaviour. Shaping accelerates learning and can change the optimal behaviour.

Sparse Reward Research

Doctoral study examines problems where feedback arrives only very rarely. Sparse feedback makes learning extremely difficult without additional guidance.

Delayed Reward Research

Research examines consequences appearing long after the responsible action. Delay makes connecting cause with outcome genuinely very difficult.

Discounting Research

Doctoral work examines weighting immediate outcomes against distant future ones. Discount choice substantially changes what behaviour appears optimal.

Return Definition Research

Research examines how accumulated future outcomes should best be summarised. Return definition determines exactly what the agent is optimising.

Value Function Research

Doctoral study examines estimating long term outcome from any given situation. Value estimation underpins the majority of learning algorithms.

Action Value Research

Research examines estimating outcome from taking one specific action. Action values permit choosing behaviour without any explicit forward planning.

Advantage Estimation Research

Doctoral work examines measuring how much better one action is than average. Advantage estimates reduce variance in policy improvement methods.

Bellman Equation Research

Research examines the recursive relationship defining optimal sequential behaviour. This relationship is the mathematical core of the whole field.

Dynamic Programming Research

Doctoral study examines exact solution methods when the environment is known. These methods provide the theoretical basis for learned approximations.

Policy Iteration Research

Research examines cycling between evaluating and improving a behaviour rule. This cyclical structure recurs throughout nearly all modern algorithms.

Value Iteration Research

Doctoral work examines repeatedly refining value estimates toward optimality. This procedure underlies a great many practical learning algorithms.

Temporal Difference Learning

Research examines learning from discrepancy between successive predictions. This idea is among the most influential in the whole research field.

Monte Carlo Method Research

Doctoral study examines learning from complete observed outcome sequences. These methods are unbiased and suffer from very high estimate variance.

Eligibility Trace Research

Research examines mechanisms assigning credit across recent decision sequences. Traces interpolate between single step and full outcome learning.

Bootstrapping Research

Doctoral work examines learning estimates partly from other current estimates. Bootstrapping accelerates learning and can cause instability.

Function Approximation Research

Research examines representing values or behaviour with learned functions. Approximation is essential wherever situations are too numerous to enumerate.

Linear Approximation Research

Doctoral study examines simple approximations permitting theoretical analysis. Linear settings yield guarantees that deep methods entirely lack.

Convergence Theory Research

Research examines whether learning procedures reach any stable final solution. Many widely used methods have no convergence guarantee whatsoever.

Stability Analysis Research

Doctoral work examines conditions under which learning remains well behaved. Instability is a persistent practical problem within deep methods.

Divergence Research

Research examines learning procedures failing catastrophically rather than converging. Particular combinations of ingredients are known to diverge reliably.

Sample Complexity Research

Doctoral study examines how much interaction learning genuinely requires. Sample demands are the principal barrier to practical real deployment.

Regret Analysis Research

Research examines cumulative shortfall relative to optimal possible behaviour. Regret bounds quantify how efficiently an algorithm actually learns.

Learning Guarantee Research

Doctoral work examines formal assurances about what learning will achieve. Guarantees typically require assumptions that applications violate.

Bandit Theory Research

Research examines the simplest setting balancing trying options against exploiting. Bandit theory provides the foundation for exploration analysis.

Contextual Bandit Research

Doctoral study examines single decisions informed by observed side information. Contextual settings underpin many deployed recommendation systems.

Exploration Theory Research

Research examines principled approaches to gathering informative experience. Exploration determines whether good behaviour is ever discovered.

Optimism Principle Research

Doctoral work examines acting as though uncertain options are favourable. Optimism drives systematic rather than random information gathering.

Posterior Sampling Research

Research examines sampling beliefs to decide which behaviour to attempt. Sampling based exploration is elegant and empirically very effective.

Information Theoretic Analysis

Doctoral study examines learning through the lens of information acquisition. This perspective unifies exploration across differing problem settings.

Representation Theory Research

Research examines what properties useful learned representations must satisfy. Theory explains why some representations support learning and others fail.

Generalisation Theory Research

Doctoral work examines behaviour transferring to situations never encountered. Generalisation theory here lags substantially behind supervised learning.

Horizon Length Research

Research examines how far ahead decisions must reasonably be considered. Long horizons make the credit assignment problem substantially harder.

Average Reward Setting

Doctoral study examines optimising ongoing performance rather than discounted totals. This formulation suits systems intended to run indefinitely.

Risk Sensitive Objective

Research examines objectives caring about outcome variability, not only averages. Risk sensitivity matters wherever rare bad outcomes are unacceptable.

Constrained Formulation Research

Doctoral work examines optimising subject to limits that must be respected. Constraints express requirements that reward alone captures poorly.

Multi Objective Formulation

Research examines problems with several genuinely competing objectives at once. Combining goals into one number conceals the trade offs being made.

Occupancy Measure Research

Doctoral study examines describing behaviour through visitation of situations. This view converts several problems into convex optimisation.

Duality Theory Research

Research examines dual formulations of sequential decision making problems. Dual views yield algorithms and bounds that the primal view obscures.

Game Theoretic Foundation

Doctoral work examines connections between learning and strategic interaction. Game theory becomes essential once several agents learn together.

Control Theory Connection

Research examines relationships with established optimal control methods. Control theory offers guarantees that learning approaches frequently lack.

Policy Gradient Research

Doctoral study examines improving behaviour by following performance gradients. Gradient methods handle continuous actions that value methods struggle with.

Actor Critic Research

Research examines pairing a behaviour rule with a learned value estimator. This pairing underpins most modern practical learning algorithms.

Natural Gradient Research

Doctoral work examines improvement directions accounting for behaviour geometry. Natural directions give more stable progress than raw gradients.

Trust Region Method Research

Research examines limiting how much behaviour changes at each learning step. Limiting change prevents the collapse that large steps can cause.

Proximal Method Research

Doctoral study examines practical approximations to constrained improvement steps. These methods became the default choice across many applications.

Deterministic Policy Research

Research examines behaviour rules producing one action without any randomness. Deterministic formulations suit continuous control very efficiently.

Entropy Regularised Method

Doctoral work examines encouraging behavioural randomness during learning. Encouraged randomness improves exploration and learning robustness.

Maximum Entropy Framework

Research examines a formulation preferring the most random adequate behaviour. This framework connects learning with probabilistic inference.

Deep Value Learning Research

Doctoral study examines neural networks estimating action values at scale. Neural estimation enabled learning directly from raw observations.

Target Network Research

Research examines slowly changing reference estimates stabilising deep learning. Slow references prevent estimates chasing their own movement.

Experience Replay Research

Doctoral work examines reusing stored past interaction during learning. Reuse improves sample efficiency and breaks harmful data correlation.

Prioritised Replay Research

Research examines selecting the most informative stored experience for reuse. Prioritisation accelerates learning and introduces estimation bias.

Distributional Value Research

Doctoral study examines learning full outcome distributions rather than averages. Distributions support risk sensitivity and improve learning stability.

Ensemble Method Research

Research examines maintaining several estimators to quantify disagreement. Disagreement provides a practical measure of epistemic uncertainty.

Overestimation Bias Research

Doctoral work examines systematic inflation of learned value estimates. Inflation arises from maximising over quantities that are noisily estimated.

Variance Reduction Research

Research examines lowering the noise in learning signal estimates. Lower variance permits larger and considerably more reliable improvement steps.

Baseline Method Research

Doctoral study examines subtracting reference quantities to reduce estimate noise. Baselines reduce variance without introducing any additional bias.

Credit Assignment Research

Research examines attributing outcomes to the decisions that actually caused them. Attribution is the central difficulty in all sequential learning.

Long Horizon Credit Research

Doctoral work examines attribution across very extended decision sequences. Long sequences defeat the methods that much shorter problems permit.

Hierarchical Learning Research

Research examines decomposing behaviour into layered levels of decision. Hierarchy shortens the effective horizon that each level must handle.

Option Framework Research

Doctoral study examines extended behaviours treated as single composite actions. Extended actions permit reasoning at a coarser temporal scale.

Skill Discovery Research

Research examines agents autonomously identifying reusable behavioural units. Discovered skills accelerate learning on subsequent related tasks.

Temporal Abstraction Research

Doctoral work examines reasoning over differing timescales simultaneously. Abstraction makes long horizon problems computationally tractable.

Goal Conditioned Learning

Research examines behaviour rules accepting a specified target as input. One learned rule can then pursue very many differing possible goals.

Hindsight Relabelling Research

Doctoral study examines treating reached outcomes as though they were intended. Relabelling extracts learning signal from apparent failures.

Curriculum Learning Research

Research examines ordering tasks from simpler toward more difficult ones. Ordering makes problems solvable that direct attempts cannot solve.

Automatic Curriculum Research

Doctoral work examines systems generating their own progression of tasks. Automatic generation removes reliance on human designed task ordering.

Curiosity Driven Exploration

Research examines agents seeking situations they cannot yet predict well. Curiosity produces exploration where external feedback is entirely absent.

Intrinsic Motivation Research

Doctoral study examines internally generated signals driving agent behaviour. Internal signals sustain learning without any external feedback.

Novelty Search Research

Research examines seeking behavioural variety rather than direct objective progress. Variety seeking escapes local optima that direct pursuit finds.

Count Based Exploration

Doctoral work examines encouraging visits to rarely encountered situations. Counting is principled and scales poorly to large observation spaces.

Parameter Space Exploration

Research examines perturbing the behaviour rule rather than individual actions. Parameter perturbation produces more coherent exploratory behaviour.

Evolutionary Method Research

Doctoral study examines population based search over behaviour parameters. Evolutionary search is simple and parallelises extremely well indeed.

Population Based Method

Research examines maintaining many agents and selecting among them. Populations support both search and configuration tuning simultaneously.

Derivative Free Optimisation

Doctoral work examines improving behaviour without computing any gradients. These methods suit objectives that are not differentiable at all.

Model Based Learning Research

Research examines learning environment dynamics and planning within them. Model based approaches typically require far less environment interaction.

World Model Research

Doctoral study examines learned internal simulations of the environment. Internal simulation permits practice without any real world interaction.

Latent Dynamics Research

Research examines predicting change within a compressed representation. Compressed prediction avoids modelling irrelevant observational detail.

Planning With Models

Doctoral work examines choosing actions by simulating possible future outcomes. Planning trades computation at decision time for less interaction.

Tree Search Research

Research examines systematically exploring branching sequences of future decisions. Tree search combined with learning achieved landmark game results.

Receding Horizon Control

Doctoral study examines planning a short sequence and repeatedly replanning. Repeated replanning corrects for accumulating model inaccuracy.

Simulated Experience Research

Research examines learning from experience generated by a learned model. Generated experience supplements limited real world interaction data.

Model Error Research

Doctoral work examines inaccuracy within learned environment dynamics. Small inaccuracies accumulate severely across long simulated sequences.

Uncertainty Aware Planning

Research examines planning that accounts for model confidence explicitly. Awareness prevents exploiting regions where models are unreliable.

Ensemble Dynamics Research

Doctoral study examines several learned models quantifying dynamics uncertainty. Model disagreement indicates where predictions cannot be trusted.

Successor Representation

Research examines representing expected future situation visitation directly. This representation separates environment structure from objectives.

Auxiliary Task Research

Doctoral work examines additional prediction problems shaping learned representations. Auxiliary problems provide signal where feedback is very sparse.

Representation Learning Research

Research examines learning useful encodings of raw environment observations. Encoding quality frequently determines overall learning success.

Contrastive Objective Research

Doctoral study examines objectives distinguishing related from unrelated observations. Contrastive learning produces representations without any labels.

Self Supervised Pretraining

Research examines preparing representations before any objective is specified. Pretraining substantially reduces interaction required afterward.

Recurrent Policy Research

Doctoral work examines behaviour rules retaining information across timesteps. Memory is essential wherever observations are genuinely incomplete.

Attention Based Policy

Research examines behaviour rules selectively focusing on relevant information. Attention handles large and variable structured observations well.

Sequence Model Approach

Doctoral study examines treating decision problems as sequence prediction. This reframing permits reuse of powerful existing sequence architectures.

Return Conditioned Modelling

Research examines predicting actions given a desired outcome level. This approach sidesteps value estimation and its instabilities entirely.

Foundation Model Integration

Doctoral work examines combining large pretrained models with decision learning. Pretrained knowledge supplies priors that interaction cannot cheaply provide.

Language Grounded Learning

Research examines connecting language with situated agent behaviour. Language provides an accessible interface for specifying agent objectives.

Instruction Following Agent

Doctoral study examines agents pursuing goals described in natural language. Language specification is far easier than engineering reward functions.

Memory Architecture Research

Research examines structures storing and retrieving information over time. Memory design determines what long range dependencies agents capture.

Continual Learning Research

Doctoral work examines agents learning across a sequence of differing tasks. Continual settings match deployment better than isolated training.

Catastrophic Forgetting Research

Research examines losing earlier capability when learning something new. Forgetting is a fundamental obstacle to genuinely continual learning.

Offline Learning Research

Doctoral study examines learning entirely from previously collected interaction. Offline learning suits settings where exploration would be unsafe.

Distribution Shift Research

Research examines mismatch between training data and encountered situations. Shift is the central difficulty in all offline learning research.

Conservative Method Research

Doctoral work examines deliberately pessimistic estimates for unseen actions. Pessimism prevents exploiting errors in poorly supported regions.

Behaviour Regularisation Research

Research examines keeping learned behaviour close to its original data source. Staying close limits errors and also limits achievable improvement.

Dataset Quality Research

Doctoral study examines how collected data properties affect learning outcomes. Data coverage matters more than sheer quantity of interaction.

Off Policy Evaluation

Research examines estimating behaviour quality without actually deploying it. Reliable estimation is prerequisite for any safe real deployment.

Importance Weighting Research

Doctoral work examines reweighting data to reflect differing behaviour rules. Weights become extremely unstable over long decision sequences.

Doubly Robust Estimation

Research examines estimators remaining valid if either component is correct. Robustness matters where all assumptions are somewhat questionable.

Counterfactual Estimation

Doctoral study examines estimating outcomes of actions never actually taken. Counterfactual reasoning connects this field with causal inference.

Constrained Action Method

Research examines restricting choices to actions well supported by data. Restriction avoids extrapolation that causes offline learning failure.

Imitation Learning Research

Doctoral work examines acquiring behaviour by observing demonstrations. Demonstration avoids specifying objectives that are hard to articulate.

Behavioural Cloning Research

Research examines directly copying demonstrated actions through supervised learning. Direct copying is simple and degrades outside demonstrated situations.

Error Accumulation Research

Doctoral study examines small mistakes leading agents progressively further astray. Accumulating deviation is the classic weakness of imitation methods.

Inverse Reinforcement Learning

Research examines recovering objectives from observed expert behaviour. Recovered objectives transfer considerably better than copied actions do.

Reward Inference Research

Doctoral work examines inferring what people actually want from their behaviour. Inference must account for human error and bounded rationality.

Preference Based Learning

Research examines learning objectives from comparisons between options. Comparisons are far easier for people to provide than numeric scores.

Human Feedback Research

Doctoral study examines incorporating human judgement into agent learning. Human feedback now underpins much practical alignment of large models.

Comparison Data Research

Research examines collecting and using human preference comparison data. Data collection design strongly shapes what objectives get learned.

Reward Model Research

Doctoral work examines learned proxies standing in for true objectives. Proxy models are imperfect and can be exploited by strong optimisation.

Reward Exploitation Research

Research examines agents achieving high scores without the intended behaviour. Exploitation reveals gaps between stated and genuine objectives.

Specification Gaming Research

Doctoral study examines agents satisfying the letter of a specification only. Gaming examples are numerous, documented and frequently instructive.

Goal Misgeneralisation Research

Research examines agents pursuing unintended goals in new situations. Capable behaviour can accompany an entirely mistaken learned objective.

Sim To Real Transfer

Doctoral work examines moving behaviour learned in simulation onto hardware. Transfer failure is the principal obstacle in robotic applications.

Domain Randomisation Research

Research examines varying simulation conditions to produce robust behaviour. Variation forces behaviour that tolerates unmodelled differences.

Transfer Learning Research

Doctoral study examines reusing learned capability across differing tasks. Transfer reduces the interaction each new task actually requires.

Meta Learning Research

Research examines learning procedures that themselves adapt very rapidly. Meta learning targets fast adaptation rather than any single solution.

Few Shot Adaptation Research

Doctoral work examines adjusting behaviour from very limited new experience. Rapid adaptation is essential for practical deployment in varied settings.

Multitask Learning Research

Research examines learning many related tasks within one shared system. Shared learning can help and can also cause harmful task interference.

Generalist Agent Research

Doctoral study examines single agents handling very diverse capability demands. Generalist systems trade peak performance for genuine breadth.

Zero Shot Generalisation

Research examines performing acceptably within entirely unseen situations. Generalisation here remains far weaker than in supervised learning.

Multi Agent Learning Research

Doctoral work examines several agents learning within a shared environment. Simultaneous learning removes the stationarity single agent theory assumes.

Cooperative Setting Research

Research examines agents sharing a common objective and coordinating action. Coordination is difficult even when their interests align perfectly.

Competitive Setting Research

Doctoral study examines agents with directly opposing objectives. Competition drives capability and complicates evaluation quite considerably.

Mixed Motive Setting

Research examines settings combining shared and conflicting interests. Mixed motives describe most genuinely realistic interaction situations.

Self Play Research

Doctoral work examines agents improving by competing against their own copies. Self play produced several landmark results within game playing.

Emergent Communication Research

Research examines agents developing signalling systems to coordinate action. Emergent signalling illuminates conditions favouring communication generally.

Team Credit Assignment

Doctoral study examines attributing shared outcomes to individual contributors. Attribution determines whether individual agents learn usefully.

Nonstationarity Research

Research examines environments changing because other agents are also learning. Moving targets defeat convergence arguments built for fixed problems.

Equilibrium Learning Research

Doctoral work examines learning procedures reaching stable strategic outcomes. Equilibrium concepts define what stability even means in this setting.

Mechanism Design Research

Research examines designing interaction rules producing desirable collective behaviour. Rule design shapes outcomes more reliably than agent design.

Social Dilemma Research

Doctoral study examines situations where individual and collective interest diverge. These settings test whether cooperation can emerge at all.

Population Dynamics Research

Research examines how strategies spread through whole populations of agents. Population views connect learning with evolutionary game theory.

Opponent Modelling Research

Doctoral work examines agents predicting how other agents will behave. Modelling others enables anticipation rather than mere later reaction.

Scalability Research

Research examines methods remaining effective as agent numbers grow very large. Many approaches fail entirely beyond a handful of participants.

Distributed Training Research

Doctoral study examines learning across many parallel computational workers. Parallel interaction collection accelerates otherwise impractical training.

Safe Learning Research

Research examines learning while avoiding unacceptable outcomes throughout. Safety during learning is required for any physical world deployment.

Constrained Policy Optimisation

Doctoral work examines improving behaviour while satisfying stated limits. Constraint satisfaction during learning is genuinely difficult to guarantee.

Safe Exploration Research

Research examines gathering information without entering dangerous situations. Safe exploration is essential wherever mistakes carry real cost.

Shielding Research

Doctoral study examines external components blocking unacceptable agent actions. Shields provide guarantees that learned behaviour alone cannot.

Formal Verification Research

Research examines mathematically proving properties of learned behaviour. Verification is difficult and increasingly demanded for critical systems.

Robustness Research

Doctoral work examines behaviour tolerating disturbance and unmodelled conditions. Robustness determines whether laboratory results survive deployment.

Adversarial Perturbation Research

Research examines behaviour under deliberately constructed harmful conditions. Learned agents are frequently brittle to small crafted disturbances.

Worst Case Analysis

Doctoral study examines performance under the least favourable circumstances. Worst case reasoning matters where average performance is insufficient.

Uncertainty Estimation Research

Research examines agents recognising the limits of their own knowledge. Recognised uncertainty supports both caution and efficient exploration.

Interpretability Research

Doctoral work examines understanding what learned agents have internally represented. Understanding is prerequisite for trusting autonomous decision systems.

Policy Explanation Research

Research examines conveying why an agent selected particular actions. Explanation supports oversight by people who did not build the system.

Human Oversight Research

Doctoral study examines people supervising and intervening in agent behaviour. Effective oversight requires agents that remain genuinely interruptible.

Interruptibility Research

Research examines agents accepting correction without learning to avoid it. Optimisation pressure can discourage accepting human intervention.

Value Alignment Research

Doctoral work examines ensuring agent objectives match human intentions. Alignment difficulty grows as agent capability increases substantially.

Scalable Oversight Research

Research examines supervising agents on tasks people cannot easily evaluate. Oversight must scale as capability exceeds unaided human judgement.

Reward Misspecification Research

Doctoral study examines objectives failing to capture what was actually wanted. Misspecification is arguably the deepest problem in this field.

Side Effect Research

Research examines unintended consequences of pursuing a stated objective. Objectives rarely specify what the agent should leave undisturbed.

Instrumental Behaviour Analysis

Doctoral work examines behaviours useful across many differing objectives. Some instrumental behaviours conflict with human oversight goals.

Deployment Risk Research

Research examines harms arising when learned agents act in the world. Deployment risk differs substantially from measured training performance.

Evaluation Methodology Research

Doctoral study examines how agent capability should properly be assessed. Evaluation practice within this field has been very widely criticised.

Benchmark Design Research

Research examines constructing tasks meaningfully testing agent capability. Benchmark choice determines what the field believes has improved.

Reproducibility Research

Doctoral work examines whether published results can be independently repeated. Reproduction here is notoriously difficult and frequently fails.

Statistical Reporting Research

Research examines how results and their variability should be reported. Too few random seeds produce conclusions that simply do not hold up.

Hyperparameter Sensitivity

Doctoral study examines performance depending heavily on configuration choices. Extreme sensitivity undermines claims about algorithmic advances.

Compute Efficiency Research

Research examines the computational resources that learning genuinely requires. Resource demands limit who can participate in this research at all.

Environment Design Research

Doctoral work examines constructing tasks that meaningfully challenge agents. Environment properties determine what capabilities can be demonstrated.

Simulation Platform Research

Research examines software environments supporting learning experiments. Platform fidelity determines how far results transfer to actual reality.

Robotics Application Research

Doctoral study examines learned control of physical robotic systems. Physical systems impose safety and sample constraints simulation lacks.

Manipulation Application

Research examines learned control of grasping and object handling tasks. Manipulation involves contact physics that remains difficult to model.

Locomotion Application

Doctoral work examines learned control of walking and legged movement. Locomotion is among the most successful physical applications so far.

Autonomous Driving Research

Research examines learning applied to vehicle decision making. Safety requirements sharply constrain what learning approaches are acceptable.

Industrial Control Application

Doctoral study examines learned control of manufacturing and process systems. Industrial settings demand reliability that learning rarely guarantees.

Energy System Application

Research examines learning applied to power grids and building systems. Energy applications offer large efficiency gains and low failure tolerance.

Healthcare Application Research

Doctoral work examines sequential decision learning within health settings. Health applications must learn offline because exploration harms patients.

Treatment Strategy Research

Research examines learning sequences of treatment decisions from records. Learned strategies require validation that observational data cannot supply.

Recommendation Application

Doctoral study examines sequential learning within content and product recommendation. Deployed systems shape the preferences they subsequently learn from.

Operations Application Research

Research examines learning applied to scheduling, routing and inventory. Operational problems have structure that general methods entirely ignore.

Finance Application Research

Doctoral work examines sequential decision learning within financial settings. Financial environments are adversarial and profoundly nonstationary.

Scientific Discovery Application

Research examines agents directing experiments and scientific search. Sequential experiment selection suits this framework particularly well.

Game Playing Research

Doctoral study examines games as testbeds for decision making research. Games provide clear objectives and unlimited cheap interaction data.

Education Application Research

Research examines sequencing learning material adaptively for students. Educational applications raise distinctive fairness and consent questions.

Ethics Research

Doctoral work examines ethical questions raised by autonomous learning agents. Agents optimising objectives affect people who never consented.

Governance Research

Research examines oversight arrangements for deployed autonomous systems. Governance determines accountability when learned agents cause harm.

Workforce And Skills Research

Doctoral study examines expertise required to build and operate these systems. Combined theoretical and engineering expertise is genuinely scarce.

Implementation And Adoption

Research examines why these methods reach practice or fail to do so. Very few published advances survive contact with real world deployment.