Date: Wednesday 16th January 2013, 11:15, CSE102/103 Computer Science, Univ. of York
Group Meeting
Date: Wednesday 23rd January 2013, 11:15, CSE102-103 Computer Science, Univ. of York
Speaker: James Cussens, Department of Computer Science, Univ. of York
Topic:
In this talk I will talk about our work (past, present and future) on applying integer linear programming (ILP) to model selection. Model selection is that statistical problem of determining the qualitative structure of some (notional) probability distribution lying behind observed data. Almost always the best we can hope for is to find the most probable model. In this talk I will be mainly focusing on the case where the set of candidate models are all Bayesian networks. However, I will also mention current/future work on Dynamic Multiregression Models, Chain Event Graphs and "structural imsets". In each case we use an ILP solver to find the most probable model.
Date: Wednesday 30th January 2013, 11:15, CSE102-103 Computer Science, Univ. of York
Speaker: Ed Powley, Department of Computer Science, Univ. of York
Topic: Information capture and reuse in Monte Carlo Tree Search
Monte Carlo Tree Search (MCTS) has produced a huge leap in AI player
strength for a range of games and decision problems. While MCTS has proved
effective in its "vanilla" form, many enhancements exist to improve
its performance. These enhancements can be seen as capturing information from
one part of the search tree and reusing it to guide the search in other parts.
In this talk, I will introduce a framework for understanding, designing and
combining these information capture and reuse enhancements. I will use this
framework to give some examples of MCTS enhancements from the literature (which
have been instrumental in the successes of MCTS to date). The framework allows
us to form a clear picture of how each enhancement shares information across the
tree, and I will discuss how this influences the choice of which enhancements to
use for a given domain. I will present a new enhancement, EPisodic
Information Capture (EPIC), which we designed within the framework
explicitly to exploit the round-based structure of many games.
Date: Wednesday 13th February 2013, 11:15, CSE102-103 Computer Science, Univ. of York
Speaker: Alan Frisch, Department of Computer Science, Univ. of York
Topic: Learning to Build Effective Constraint Models - URO (Unidentified Research Object)
CONJURE is a rule-based system that can transform an abstract description of a combinatorial problem into a set of alternative constraint models, each of which can be used by a constraint solver to generate solutions to the original problem. The alternative constraint models typically vary greatly in terms of the computational resources each requires to solve the problem. CONJURE currently is unable to identify which of the models it generates are effective; the only guarantee is that the models are correct.
I will present some prelimary ideas on how CONJURE could be endowed with the ability to automatically learn to generate only those constraint models that are likely to be among the most effective at solving a given problem. Reinforcement learning, Monte Carlo tree search and local search are among the techniques that could prove useful.
A good deal of time will be devoted to discussion and I look forward to hearing everyone's suggestions on how best to take this work forward.
Note: It does not matter if you know nothing about CONJURE or constraint modelling. The issues will be discussed at a level that applies to almost any rule-based system, so the details of CONJURE are largely irrelevant.
Date: Wednesday 20th February 2013, 11:15, CSE102-103 Computer Science, Univ. of York
Speaker: Garo Panikian, Department of Computer Science, Univ. of York
Topic: Statistical inference of dynamical systems with application to modelling fish population
We test the hypothesis whether a heteroscedastic (non-constant variance) model is a good choice for describing the relationships between survival variability and adult abundance of fish population. This relationship is affected by many regulatory factors such as: food, salinity, water temperature and so forth, which would all behave as stochastic processes. Our tests use statistical procedures applied on 257 stock-recruitment time series datasets obtained from Ransom Myers' database where we fit systematically a heteroscedastic model for each dataset respectively. We analyse the shape of the likelihood function, select an appropriate optimisation technique to determine whether we can improve the power of our statistical analysis. We estimate the confidence intervals by simulating 1,000 new datasets and fit the heteroscedastic model to obtain parameter estimates. We then use the probability mass function to determine whether we can recover the sign of the coefficient of heteroscedasticity accurately. To validate our results, we repeat the analysis using a Bayesian framework where we define a flat prior over the parameters and then estimate the marginal posterior using Hamiltonian Monte Carlo.
We conclude that heteroscedastic models are case by case dependent and applicable only if the large portion of the probability mass function of the coefficient of heteroscedasticity falls in the negative region.
Date: Wednesday 27th February 2013, 11:15, CSE083 Computer Science, Univ. of York
Group Meeting
Date: Wednesday 6th March 2013, 11:15, CSE083 Computer Science, Univ. of York
Speaker: Pete Scopes, Department of Computer Science, Univ. of York
Topic: ReAdaptive Tile-Coding (work in progress)
Reinforcement Learning (RL) is a simple but effective means of problem-solving; using the basic learning concept of trial and error an agent can solve complex problems. Some problems have continuous state spaces that make vanilla RL techniques stubmle since tere are a theoretical infinite number of states. Such problems can have a natural "resolution" which brings the number of states down to a finite number. Function Approximation (FA) aims to make use of this directly or create its own resolution to find a "good" (sub-optimal) solution. Tile-Coding is a popular FA algorithm which can suffer from deciding between speed and quality of learning; Adaptive Tile-Coding bypasses this problem by going from a coarse to a fine grain view of the problem in an intelligent manor.
I will introduce the concepts of FA and Tile-Coding, and present a further, promising, advancement on Adaptive Tile-Coding called ReAdaptive Tile-Coding.
Date: Wednesday 13th March 2013, 11:15, CSE083 Computer Science, Univ. of York
Speaker: Mohd Farid Md Alias, Department of Computer Science, Univ. of York
Topic: AMIR-III Robotic Head and Book Shelving System
Advances in humanoid technology have improved the capacity and capability of humanoids to perform human tasks within human-existing environments. Such situation would demand humanoids to always operate in a safe and friendly way to incite general acceptance towards humanoids. Thus there is a need to develop human-friendly humanoids particularly with expressive face to portray emotions. These robots, known as humanoid heads, can be of anthropomorphic (human-like) or iconic (cartoonish) look. Iconic humanoid heads are generally less prone to anthropomorphism pitfalls of Mori’s Uncanny Valley theory. In this research, a new humanoid head design with unique mouth mechanism is developed. The humanoid head, named as AMIR-III is capable of producing five basic facial expressions based on AMIR Model of Facial Expression (AMEr). Library users often found the problem of finding their desired books on the designated shelves due to misplaced. For a library with RFID book tagging system, such issue could be tacked through regular shelf scanning by librarians to identify and sort any misplaced books. However, such task would be time consuming and mundane to human librarians. This research proposes an automated robotic system to execute such task and connect its scanning results back to the library database system.
Speaker: Alexandros Komninos, Department of Computer Science, Univ. of York
Topic: An Application for Entity Ranking on the Web
In the context of the presented application, the focus is on named entities of the three general categories: person, location, organization. Entity ranking is the process of returning a sorted list of entities according to relevance in response to a query, for example "British Nobel Prize winners" or "countries where you can pay with Euro". The application takes a collection of documents returned by a web search engine, uses a named entity recognizer to extract the entities and ranks them using information retrieval features (frequency, document frequency, ranking of source document). Several ranking models based on the above quantities were tested, along with algorithms for estimating the amount of relevant results from the distribution of scores.
Date: Wednesday 24th April 2013, 11:15, CSE083 Computer Science, Univ. of York
Speaker: Keshav Dahal, Department of Computer Science, Bradford University
Topic: Acquaintance-based trust model for the evolution of cooperation in distributed systems
In this talk, I will present a possible improvement to a reputation model of online market places, like that of eBay, with our interest lying on investigating how the cooperativeness and population of co-operators would evolve if the weight of the feedback source was assigned on basis of past association between players. The feedback sources are categorised into different types to define an aggregation method for trustworthiness asssessment that considers applying a dynamically computed weight to each source of feedback. A genetic algorithm based spatial iterated prisoner's dilemma (SIPD) environment has been used to simulate the experiments to study the evolution of cooperativeness. Our simulation results show that breaking feedback sources on the basis of acquaintance and assigning weight accordingly favour the evolution of cooperativeness in the player society when compared to models which do not classify the feedback sources.
Date: Wednesday 1st May 2013, 11:15, CSE083 Computer Science, Univ. of York
Topic: Group Meeting
Date: Thursday 25th April 2013, 15:00, CSE203 Computer Science, Univ. of York
Speaker: Akihiro Katsura, Department of Computer Science, Univ. of York
Topic: Question Answering in the Generation of Social Networking [MSc by Research Literature Review seminar]
The performance of Question Answering systems is enhanced by developing Machine Learning approaches while they requires a large amount of training data. A reasonable means to come over the obstacle may be to utilise numerous instances in Social Networking Services, especially Social QA websites like Yahoo! Answers for QA systems. In this review, I will introduce how QA systems were developed and how Social QA websites were utilised; and will discuss problems for further development of QA systems.
Date: Wednesday 8th May 2013, 11:15, CSE083 Computer Science, Univ. of York
Speaker: Kiran Frenandes
Topic: Gun crime hotspot modeling using systems-based approach
Rise in gun crime incidents are major cause of concern for the UK legal system. The Government and police are trying hardest to curb this growing threat. To support the traditional crime analysis, with the advancement in technology, artificial expert systems are being incorporated into police services. This project was an active collaboration between the UK police and The York Centre for Complex Systems Analysis, The University of York, UK, aim of which was to aid the police department with a set of 'tools' in predicting and preventing gun crime which is not fully touched upon yet. A software system was developed to predict the hot spots of gun crime based upon different measurable features. The contribution of this paper also lies in the demonstration of how systems-based approaches may be used to support representations of social problems from an operations perspective.
Date: Thursday 9th May 2013, 11:30, CSE083 Computer Science, Univ. of York
Speaker: Farid Alias
Topic: Reinforcement Learning for Robotics [PhD Literature Review]
Reinforcement learning is a computational approach to learning by an agent from direct interactions with its environment so as to maximise its accumulated rewards. Dynamic Programming, Monte Carlo methods and Temporal-Difference Learning are the three fundamental classes of reinforcement learning techniques; each with its own strengths and weaknesses. Robotics is an interesting platform where reinforcement learning can be potentially applied particularly on sequential behaviour generation problems for autonomous robots. Nevertheless, the issues of dimensionality, real world samples, real world interactions and goal specification are the major challenges to be faced in implementing reinforcement learning in robotics. Two robotics examples which apply advanced reinforcement learning techniques are described namely robot walking gait and autonomous helicopter flight.
Date: Wednesday 15th May 2013, 11:15, CSE083 Computer Science, Univ. of York
Speaker: Peter Gregory, School of Computing, Teesside University
Topic: Planning Modulo Theories: Extending the Planning Paradigm
Considerable effort has been spent extending the scope of planning beyond propositional domains to include, for example, time and numbers. Each extension has been designed as a separate specific semantic enrichment of the underlying planning model, with its own syntax and customised integration into a planning algorithm. Inspired by work on SAT Modulo Theories (SMT) in the SAT community, we develop a modelling language and planner that treat arbitrary first order theories as parameters. We call the approach Planning Modulo Theories (PMT). We introduce a modular language to represent PMT problems and demonstrate its benefits over PDDL in expressivity and compactness. We present a generalisation of the h_max heuristic that allows our planner, PMTPlan, to automatically reason about arbitrary theories added as modules. Over several new and existing benchmarks, exploiting different theories, we show that PMTPlan can significantly out-perform an existing planner using PDDL models.
Date: Wednesday 29th May 2013, 11:15, CSE083 Computer Science, Univ. of York
Speaker: Daniel Whitehouse, Department of Computer Science, Univ. of York
Topic: Integrating Monte Carlo Tree Search with Knowledge-Based Methods to Create Engaging Play in a Commercial Mobile Game
Monte Carlo Tree Search (MCTS) has produced many recent breaktroughs in game AI research, particularly in computer Go. In this talk I will explain how MCTS can be applied to create engaging AI for a popular commercial mobile phone game: Spades by AI Factory, which has been downloaded more than 2.5 million times. In particular, I will describe how MCTS can be integrated with knowledge-based methods to create an interesting, fun and strong player which makes far fewer blunders than MCTS without injection of knowledge. These blunders are particularly noticeable for Spades, where a human player must co-operate with an AI partner. MCTS gives objectively stronger play than the knowledge-based approach used in earlier versions of the game and offers the flexibility to customise behaviour whilst maintaining a reusable core, with a reduced development cycle compared to purely knowledge-based techniques.
Date: Wednesday 5th June 2013, 11:15, CSE083 Computer Science, Univ. of York
Speaker: Sam Devlin, Department of Computer Science, Univ. of York
Topic: Potential-Based Reward Shaping for Partially Observable Environments and Multi-Agent Systems with Sparse Interactions
In this talk I will cover two recent collaborative publications that myself and Daniel have been working on. The first, with Adam Eck and Professor Leen-Kiat Soh from the University of Nebraska-Lincoln, applies potential-based reward shaping to online planning for active sensing. This work includes two novel theoretical results explaining the implications of partial observability and finite episodes on potential-based reward shaping. The second, with Yann-Michaël De Hauwere and Ann Nowé from Vrije Universiteit Brussel, introduces context sensitive reward shaping; a novel design applying different potential functions to different situations.
Speaker: David Zendle, Department of Computer Science, Univ. of York
Topic: Authoring Interactive Narrative Content for Video Games
The procedural generation of narratives with which users can interact (Interactive Narrative) has been the subject of AI research for more than a decade. Within this time-span, however, very few implementations have been produced which are complex enough or big enough to create truly enjoyable experiences or foster engagement outside of the academic community. This lack of concrete results can be ascribed to what is termed the 'authoring bottleneck'; the exponential relationship between the interactivity or size of a narrative and the amount of content which needs to be authored to implement it. In this talk I will describe the various methodologies and tools which researchers have suggested to overcome this bottleneck and also discuss the limitations of each technique. I will then go on to describe how we can begin to develop an authoring solution for Interactive Narrative within the specific domain of video games.
Date: Wednesday 12th June 2013, 11:15, CSE083 Computer Science, Univ. of York
Speaker: Peter Cowling, Department of Computer Science, Univ. of York
Topic: Departmental Seminar
Date: Wednesday 19th June 2013, 11:15, CSE083 Computer Science, Univ. of York
Speaker: Kleanthis Malialis, Department of Computer Science, Univ. of York
Topic: Multiagent Router Throttling: Decentralized Coordinated Response against DDoS Attacks
Distributed denial of service (DDoS) attacks constitute a rapidly evolving threat in the current Internet. In this paper we introduce Multiagent Router Throttling, a decentralized DDoS response mechanism in which a set of upstream routers independently learn to throttle traffic towards a victim server. We compare our approach against a baseline and a popular throttling technique from the literature, and we show that our proposed approach is more secure, reliable and cost-effective. Furthermore, our approach outperforms the baseline technique and either outperforms or has the same performance as the popular one.
Date: Wednesday 26th June 2013, 11:15, CSE083 Computer Science, Univ. of York
Group Meeting
Date: Wednesday 2nd October 2013, 11:30, CSE102-103 Computer Science, Univ. of York
Group Meeting
Date: Wednesday 16th October 2013, 11:30, CSE102-103 Computer Science, Univ. of York
Speaker: Tommy Yuan, Department of Computer Science, Univ. of York
Topic: ProtOCL: Specifying dialogue games using UML and OCL
Dialogue games are becoming increasingly popular tools for Human-Computer Dialogue and Agent Communication. However, whilst there is an increasing body of theoretical underpinning that demonstrates the value and utility of dialogue games, and also a range of novel implementations within specific problem domains, there remain very few tools to support the deployment of dialogue games based solutions within new problem domains. This talk will introduce a new approach, called ProtOCL, to the specification of dialogue games. This approach adopts Unified Modelling Language (UML) and the Object Constraint Language (OCL) and enables the rapid movement from specification to deployment and execution. This approach goes beyond existing description languages and their supporting tools by (1) using a description language that is familiar to a far larger user group, and, (2) enables code-generation using languages and technologies that are current industry standards.
Date: Wednesday 30th October 2013, 11:30, CSE102-103 Computer Science, Univ. of York
Speaker: Daniel Whitehouse, Univ. of York
Topic: Thesis Seminar - Exploiting Monte Carlo Simulations for Decision Making in Imperfect Information Games
Since October 2010 I have been part of a three year EPSRC funded project entitled "UCT for Games and Beyond" in partnership with the University of Essex and Goldsmiths College. In this talk I will present an overview of my work on Monte-Carlo Tree Search (MCTS) applied to Imperfect Information games, which forms the basis of my PhD thesis. MCTS is a family of game tree search algorithms which are behind recent advances in the fields computer Go and general game playing and increasingly, used in commercial applications. My work has focused on games with hidden information. Until the development of MCTS, previous search techniques for Imperfect Information games either required vast computational resources to solve small problems or used unsound simplifications and extensive domain knowledge to make decisions quickly enough for commercial games.
A key aspect of MCTS is the use of random simulated games to inform decision making. I will present my work on learning effective opponent models from simulated games within the MCTS framework, which enables good forward planning in large complex games with imperfect information. In particular I will present an overview of the Information Set MCTS (ISMCTS) family of algorithms, which is the key contribution of my research. The ISMCTS algorithm can effectively handle different sources of imperfect information and can produce good decisions on modest computational budgets. I will discuss the lessons learned from our recent work deploying ISMCTS into a commercial product, Spades by AIFactory.
Week 7: Monday 11th November 2013, 10:00 - 10:45, CSE082 Computer Science, Univ. of York
Speaker: Philip Mourdjis, Univ. of York
Topic: EngD Literature Review Seminar
Week 7: Monday 11th November 2013, 10:45 - 11:30, CSE082 Computer Science, Univ. of York
Speaker: Yujie Chen, Univ. of York
Topic: EngD Literature Review Seminar
Week 7: Monday 11th November 2013, 14:00 - 14:45, CSE082 Computer Science, Univ. of York
Speaker: Nick Sephton, Univ. of York
Topic: EngD Literature Review Seminar
Week 7: Wednesday 13th November 2013, 11:30, CSE102-103 Computer Science, Univ. of York
Speaker: Daniel Neagu, Department of Computing, Univ. of Bradford
Topic: A framework for comparing heterogeneous objects: on the similarity measurements for fuzzy, numerical and categorical attributes
Real-world data collections are often heterogeneous (represented by a set of mixed attributes data types: numerical, categorical and fuzzy); since most available similarity measures can only be applied to one type of data, it becomes essential to construct an appropriate similarity measure for comparing such complex data.
In this presentation, a framework of new and unified similarity measures is introduced for comparing heterogeneous objects described by numerical, categorical and fuzzy attributes. Some practical examples will illustrate, compare and discuss the applications and efficiency of the proposed approach to heterogeneous data comparison and clustering.
Week 8: Wednesday 20th November 2013, 11:30, CSE102-103 Computer Science, Univ. of York
Speaker: Dr. Julian F. Miller, Department of Electronics, Univ. of York
Topic: Cartesian Genetic Programming: An automated problem solving technique
Cartesian Genetic Programming (CGP) is a form of automatic program induction that uses an evolutionary algorithm to evolve graph-based representations of computational structures. It is a highly flexible and general technique that can find solutions in many problem domains (e.g. neural networks, mathematical equation induction, object recognition in images, function optimization, digital and analogue circuit design, combinatorial optimization, etc.). Since its invention in 1999, it has been developed and made more efficient in various ways. It can automatically capture and evolve sub-functions (known as modules) and through the introduction of self-modification operators it is possible to find mathematically provable general solutions to classes of problems. This talk is given by the inventor of the technique. The first edited book on CGP was published by Springer in September 2011. For further information CGP has its own dedicated website: http://www.cartesiangp.co.uk
Week 9: Wednesday 27th November 2013, 11:30, CSE102-103 Computer Science, Univ. of York
Speaker: Dong Li, The York Management School, Univ. of York
Topic: An Approximate Dynamic Programming Approach to the Car Rental Revenue Management with the Flexible Capacity
Different from airlines or hotels, car rental companies have the flexibility to adjust the local capacity (number of cars) at their rental stations. Cars are transshipped from one station to another to achieve a better supply/demand match on a daily basis. Such flexibility is often called shuttling in practice. This work studies the car rental revenue management with flexible capacity due to shuttling. The problem is modelled as a discrete Markov Decision Process, the exact solution to which is unfortunately not possible for any practical scenarios (the curse of dimensionality). We approximate the value function by a method of decomposing the multiple length of rentals. Numerical studies show that the approximation is a tight upper bound to the original problem. We calculate a lower bound using a heuristic policy that is constructed directly from solutions of the decomposed problem.
Week 10: Wednesday 4th December 2013, 11:30, CSE102-103 Computer Science, Univ. of York
Topic: Reserved for Literature Review Seminars
Week 11: Tuesday 10th December 2013, 10:00, CSE/082.
Speaker: Nils Monning, Department of Computer Science, Univ. of York
Topic: Approaches To Compositional Distributional Semantics
For various tasks in Information Retrieval and Natural Language Processing it is crucial to assess the meaning of words, phrases and sentences. In the attempt to find a semantic representation of complex expressions the idea for Compositional Distributional Semantics arose. It combines two well-studied paradigms: Compositionality (Frege 1879) and the Distributional Hypothesis (Harris 1954). This literature review aims to give an overview and comparison of approaches to Compositional Distributional Semantics. Further it emphasises the use of quantum theoretic features to model complex semantics.