Information audit

The information audit (IA) extends the concept of auditing from a traditional scope of accounting and finance to the organisational information management system. Information is representative of a resource which requires effective management and this led to the development of interest in the use of an IA. Prior the 1990s and the methodologies of Orna, Henczel, Wood, Buchanan and Gibb, IA approaches and methodologies focused mainly upon an identification of formal information resources (IR). Later approaches included an organisational analysis and the mapping of the information flow. This gave context to analysis within an organisation's information systems and a holistic view of their IR and as such could contribute to the development of the information systems architecture (ISA). In recent years the IA has been overlooked in favour of the systems development process which can be less expensive than the IA, yet more heavily technically focused, project specific (not holistic) and does not favour the top-down analysis of the IA. == Definition == A definition for the Information Audit cannot be universally agreed-upon amongst scholars, however the definition offered by ASLIB received positive support from a few notable scholars including Henczel, Orna and Wood; “(the IA is a) systematic examination of information use, resources and flows, with a verification by reference to both people and existing documents, in order to establish the extent to which they are contributing to an organisation’s objectives” In summary, the term audit itself implies a counting, the IA being much the same yet it counts IR and analyses how they are used and how critical they are to the success of a given task. == Role and scope of an IA == In much the same way as the IA is difficult to define, it can be utilised in a range of contexts by the information professional, from complying with freedom of information legislation to identifying any existing gaps, duplications, bottlenecks or other inefficiencies in information flows and to understand how existing channels can be used for knowledge transfer In 2007 Buchanan and Gibb developed upon their 1998 examination of the IA process by outlining a summary of its main objectives: To identify an organisation’s information resource To identify an organisation’s information needs Furthermore, Buchanan and Gibb went on to state that the IA also had to meet the following additional objectives: To identify the cost/benefits of information resources To identify the opportunities to use the information resources for strategic competitive advantage To integrate IT investment with strategic business initiatives To identify information flow and processes To develop an integrated information strategy and/or policy To create an awareness of the importance of Information Resource Management (IRM) To monitor/evaluate conformance to information related standards, legislations, policy and guidelines. == Methodology evolution == === Overview === In 1976 Riley first published a definition of IA as a way of analysing IR based on a cost-benefit model. Since Riley, scholars have outlined further developed methodologies. Henderson took a cost-benefit approach hoping to draw focus from manpower-costing to information storage and acquisition which he felt was being overlooked. In 1985 Gillman focused upon identifying the relationships which existed between various components in order to map them to one another. Neither Henderson nor Gillman’s methods offered alternative approaches beyond the existing organisational frameworks. Quinn took a hybrid-approach combining Gillman and Henderson’s methods to identify the purpose of existing IR and to position them within the organisation, as did Worlock. The differentiator between Quinn and Worlock lay in Worlock’s consideration of solutions outside of the current organisational structure. These approaches had thus far had paid little attention to the needs of the user or in making structured recommendations for the development of a corporate information strategy. Therefore, here follows a brief outline and overall comparison of four published strategic approaches in order that one might understand the development of the IA methodology. === Burk and Horton === In 1988 Burk and Horton developed InfoMap, the first IA methodology developed for widespread use. It aimed to discover, map and evaluate the IR within an organisation using a 4-stage process: Survey staff using questionnaires/interviews Measure the IR against cost/value Analyse resources Synthesise the findings and map the strengths and weaknesses of the IR against the objectives of the organisation. Although the method inventoried all IR (and therefore met standard ISO 1779) this bottom-up approach revealed limited analysis of the organisation holistically and the steps were not explicit enough. === Orna === Orna produced a top-down methodology in contrast to Burk and Horton, placing emphasis upon the importance of organisational analysis and aimed to assist in the production of a corporate information policy. Initially the method had just 4-stages, this later revised to a 10-stage process which included pre and post-audit stages as below: Conduct a preliminary review to confirm operational/strategic direction Gain support/resource from management Gain commitment from the other stakeholders (staff) Planning including the project, team, tools and techniques Identify the IR, information flow and produce a cost/value assessment Interpret findings based upon current versus desired state Produce a report to present findings Implement recommendations Monitor effects of change Repeat the IA Orna’s method introduced the need for a cyclical IA to be put in place in order for the IR to be continually tracked and improvements made regularly. Again this method was criticised for lacking some practical application and in 2004 Orna revised the methodology once more to try to rectify this problem === Buchanan and Gibb === In 1998, similarly to Orna's earlier publication, Buchanan and Gibb took a top-down approach, drawing techniques from established management disciplines to provide a framework and a level of familiarity for information professionals. This set of techniques was a notable contribution to IA methodologies and understood the need to be flexible for each organisation. Theirs was a 5-stage process: Promote benefits of the IA through seminars/surveys/CEO letter for cooperation Identify the mission objectives of the organisation, define environment (PEST), map information flow and examine organisation culture. Analyse and formulate action plan for problem areas, flow diagrams and a report of findings and recommendations Account for cost of IR and related services using Activity Based Costing (ABC) and Output Based Specification (OBS). Synthesise the whole process in final audit report and provide an information strategy (strategic direction) in relation to the organisation’s mission statement. This was the introduction of a new approach to costing the IR and had an integrated strategic direction, yet the scholars admitted that this method may be impractical for smaller organisations. === Henczel === Henczel’s methodology drew upon the strengths of Orna and Buchanan and Gibb to produce a 7-stage process: Planning and submission of business case for approval to proceed Data collection and development of an IR database and population through survey techniques Structured data analysis Data evaluation, interpretation and formulation of recommendations Communication of recommendations through a report Implementing recommendations through a devised programme The IA as a continuum-establishment of a cyclical process Focus was made once more on the strategic direction of the organisation conducting the IA. Furthermore, Henczel made examination into the use of the IA as a first-step in the development of a knowledge audit or knowledge management strategy as discussed in the later section. == Case studies == Scholars and information professionals have since tested the above methodologies with varied results. An early case study produced by Soy and Bustelo in a Spanish financial institution in 1999 aimed to identify the use of information resources for qualitative and quantitative data analysis due to the rapid expansion of the organisation within a six-year period. Although the methodology was not explicitly credited to any of the above-mentioned scholars, it did follow a strategic (post 1990's) IA process including gaining support from management, the use of questionnaires for data collection, analysis and evaluation of the data, identification and mapping of the IR, cost-analysis and outlining recommendations to assist with the establishment of an Information policy. In addition the IA report suggested that the process would need to be continual (cyclical as Orna, Henczel and Buchanan and Gibb suggest). Conclusions of this case-study stated that th

Non-photorealistic rendering

Non-photorealistic rendering (NPR) is an area of computer graphics that focuses on enabling a wide variety of expressive styles for digital art, in contrast to traditional computer graphics, which focuses on photorealism. NPR is inspired by other artistic modes such as painting, drawing, technical illustration, and animated cartoons. NPR has appeared in movies and video games in the form of cel-shaded animation (also known as "toon" shading) as well as in scientific visualization, architectural illustration and experimental animation. == History and criticism of the term == The term non-photorealistic rendering is believed to have been coined by the SIGGRAPH 1990 papers committee, who held a session entitled "Non Photo Realistic Rendering". The term has received some criticism: The term "photorealism" has different meanings for graphics researchers (see "photorealistic rendering") and artists. For artists—who are the target consumers of NPR techniques—it refers to a school of painting that focuses on reproducing the effect of a camera lens, with all the distortion and hyper-reflections that it creates. For graphics researchers, however, it refers to an image that is visually indistinguishable from reality. In fact, graphics researchers lump the kinds of visual distortions that are used by photorealist painters into "non-photorealism". Describing something by what it is not is problematic. Equivalent (made-up) comparisons might be "non-elephant biology" or "non-geometric mathematics". NPR researchers have stated that they expect the term will disappear eventually and be replaced by the now more general term "computer graphics", with "photorealistic graphics" being the term used to describe "traditional" computer graphics. Many techniques that are used to create 'non-photorealistic' images are not rendering techniques. They are modelling techniques, or post-processing techniques. While the latter are coming to be known as 'image-based rendering', sketch-based modelling techniques, cannot technically be included under this heading, which is very inconvenient for conference organisers. The first conference on non-photorealistic animation and rendering included a discussion of possible alternative names. Among those suggested were "expressive graphics", "artistic rendering", "non-realistic graphics", "art-based rendering", and "psychographics". All of these terms have been used in various research papers on the topic, but the "non-photorealistic" term seems to have nonetheless taken hold. The first technical meeting dedicated to NPR was the ACM-sponsored Symposium on Non-Photorealistic Rendering and Animation(NPAR) in 2000. NPAR is traditionally co-located with the Annecy Animated Film Festival, running on even numbered years. From 2007 onward, NPAR began to also run on odd-numbered years, co-located with ACM SIGGRAPH. == 3D == Three-dimensional NPR is the style that is most commonly seen in video games and movies. The output from this technique is almost always a 3D model that has been modified from the original input model to portray a new artistic style. In many cases, the geometry of the model is identical to the original geometry, and only the material applied to the surface is modified. With increased availability of programmable GPU's, shaders have allowed NPR effects to be applied to the rasterised image that is to be displayed to the screen. The majority of NPR techniques applied to 3D geometry are intended to make the scene appear two-dimensional. NPR techniques for 3D images include cel shading and Gooch shading. Many methods can be used to draw stylized outlines and strokes from 3D models, including occluding contours and Suggestive contours. For enhanced legibility, the most useful technical illustrations for technical communication are not necessarily photorealistic. Non-photorealistic renderings, such as exploded view diagrams, greatly assist in showing placement of parts in a complex system. Cartoon rendering, also called cel shading or toon shading, is a non-photorealistic rendering technique used to give 3D computer graphics a flat, cartoon-like appearance. Its defining feature is the use of distinct shading colors rather than smooth gradients, producing a look reminiscent of comic books or animated films. This technique is often used to blend 3D objects and environments with 2D hand-animated elements while maintaining a consistent look. Treasure Planet movie by Disney is an example of blending these techniques. == 2D == The input to a two dimensional NPR system is typically an image or video. The output is a typically an artistic rendering of that input imagery (for example in a watercolor, painterly or sketched style) although some 2D NPR serves non-artistic purposes e.g. data visualization. The artistic rendering of images and video (often referred to as image stylization) traditionally focused upon heuristic algorithms that seek to simulate the placement of brush strokes on a digital canvas. Arguably, the earliest example of 2D NPR is Paul Haeberli's 'Paint by Numbers' at SIGGRAPH 1990. This (and similar interactive techniques) provide the user with a canvas that they can "paint" on using the cursor — as the user paints, a stylized version of the image is revealed on the canvas. This is especially useful for people who want to simulate different sizes of brush strokes according to different areas of the image. Subsequently, basic image processing operations using gradient operators or statistical moments were used to automate this process and minimize user interaction in the late nineties (although artistic control remains with the user via setting parameters of the algorithms). This automation enabled practical application of 2D NPR to video, for the first time in the living paintings of the movie What Dreams May Come (1998). More sophisticated image abstractions techniques were developed in the early 2000s harnessing computer vision operators e.g. image salience, or segmentation operators to drive stroke placement. Around this time, machine learning began to influence image stylization algorithms notably image analogy that could learn to mimic the style of an existing artwork. The advent of deep learning has re-kindled activity in image stylization, notably with neural style transfer (NST) algorithms that can mimic a wide gamut of artistic styles from single visual examples. These algorithms underpin mobile apps capable of the same e.g. Prisma In addition to the above stylization methods, a related class of techniques in 2D NPR address the simulation of artistic media. These methods include simulating the diffusion of ink through different kinds of paper, and also of pigments through water for simulation of watercolor. == Artistic rendering == Artistic rendering is the application of visual art styles to rendering. For photorealistic rendering styles, the emphasis is on accurate reproduction of light-and-shadow and the surface properties of the depicted objects, composition, or other more generic qualities. When the emphasis is on unique interpretive rendering styles, visual information is interpreted by the artist and displayed accordingly using the chosen art medium and level of abstraction in abstract art. In computer graphics, interpretive rendering styles are known as non-photorealistic rendering styles, but may be used to simplify technical illustrations. Rendering styles that combine photorealism with non-photorealism are known as hyperrealistic rendering styles. == Notable films and games == This section lists some seminal uses of NPR techniques in films, games and software. See cel-shaded animation for a list of uses of toon-shading in games and movies.

Markov model

In probability theory, a Markov model is a stochastic model used to model pseudo-randomly changing systems. It is assumed that future states depend only on the current state, not on the events that occurred before it (that is, it assumes the Markov property). Generally, this assumption enables reasoning and computation with the model that would otherwise be intractable. For this reason, in the fields of predictive modelling and probabilistic forecasting, it is desirable for a given model to exhibit the Markov property. == Introduction == Andrey Andreyevich Markov (14 June 1856 – 20 July 1922) was a Russian mathematician best known for his work on stochastic processes. A primary subject of his research later became known as the Markov chain. There are four common Markov models used in different situations, depending on whether every sequential state is observable or not, and whether the system is to be adjusted on the basis of observations made: == Markov chain == The simplest Markov model is the Markov chain. It models the state of a system with a random variable that changes through time. In this context, the Markov property indicates that the distribution for this variable depends only on the distribution of a previous state. An example use of a Markov chain is Markov chain Monte Carlo, which uses the Markov property to prove that a particular method for performing a random walk will sample from the joint distribution. == Hidden Markov model == A hidden Markov model is a Markov chain for which the state is only partially observable or noisily observable. In other words, observations are related to the state of the system, but they are typically insufficient to precisely determine the state. Several well-known algorithms for hidden Markov models exist. For example, given a sequence of observations, the Viterbi algorithm will compute the most-likely corresponding sequence of states, the forward algorithm will compute the probability of the sequence of observations, and the Baum–Welch algorithm will estimate the starting probabilities, the transition function, and the observation function of a hidden Markov model. One common use is for speech recognition, where the observed data is the speech audio waveform and the hidden state is the spoken text. In this example, the Viterbi algorithm finds the most likely sequence of spoken words given the speech audio. == Markov decision process == A Markov decision process is a Markov chain in which state transitions depend on the current state and an action vector that is applied to the system. Typically, a Markov decision process is used to compute a policy of actions that will maximize some utility with respect to expected rewards. == Partially observable Markov decision process == A partially observable Markov decision process (POMDP) is a Markov decision process in which the state of the system is only partially observed. POMDPs are known to be NP complete, but recent approximation techniques have made them useful for a variety of applications, such as controlling simple agents or robots. == Markov random field == A Markov random field, or Markov network, may be considered to be a generalization of a Markov chain in multiple dimensions. In a Markov chain, state depends only on the previous state in time, whereas in a Markov random field, each state depends on its neighbors in any of multiple directions. A Markov random field may be visualized as a field or graph of random variables, where the distribution of each random variable depends on the neighboring variables with which it is connected. More specifically, the joint distribution for any random variable in the graph can be computed as the product of the "clique potentials" of all the cliques in the graph that contain that random variable. Modeling a problem as a Markov random field is useful because it implies that the joint distributions at each vertex in the graph may be computed in this manner. == Hierarchical Markov models == Hierarchical Markov models can be applied to categorize human behavior at various levels of abstraction. For example, a series of simple observations, such as a person's location in a room, can be interpreted to determine more complex information, such as in what task or activity the person is performing. Two kinds of Hierarchical Markov Models are the Hierarchical hidden Markov model and the Abstract Hidden Markov Model. Both have been used for behavior recognition and certain conditional independence properties between different levels of abstraction in the model allow for faster learning and inference. == Tolerant Markov model == A Tolerant Markov model (TMM) is a probabilistic-algorithmic Markov chain model. It assigns the probabilities according to a conditioning context that considers the last symbol, from the sequence to occur, as the most probable instead of the true occurring symbol. A TMM can model three different natures: substitutions, additions or deletions. Successful applications have been efficiently implemented in DNA sequences compression. == Markov-chain forecasting models == Markov-chains have been used as a forecasting methods for several topics, for example price trends, wind power and solar irradiance. The Markov-chain forecasting models utilize a variety of different settings, from discretizing the time-series to hidden Markov-models combined with wavelets and the Markov-chain mixture distribution model (MCM).

Rule-based machine learning

Rule-based machine learning (RBML) is a term in computer science intended to encompass any machine learning method that identifies, learns, or evolves 'rules' to store, manipulate or apply. The defining characteristic of a rule-based machine learner is the identification and utilization of a set of relational rules that collectively represent the knowledge captured by the system. Rule-based machine learning approaches include learning classifier systems, association rule learning, artificial immune systems, and any other method that relies on a set of rules, each covering contextual knowledge. While rule-based machine learning is conceptually a type of rule-based system, it is distinct from traditional rule-based systems, which are often hand-crafted, and other rule-based decision makers. This is because rule-based machine learning applies some form of learning algorithm such as Rough sets theory to identify and minimise the set of features and to automatically identify useful rules, rather than a human needing to apply prior domain knowledge to manually construct rules and curate a rule set. == Rules == Rules typically take the form of an '{IF:THEN} expression', (e.g. {IF 'condition' THEN 'result'}, or as a more specific example, {IF 'red' AND 'octagon' THEN 'stop-sign}). An individual rule is not in itself a model, since the rule is only applicable when its condition is satisfied. Therefore rule-based machine learning methods typically comprise a set of rules, or knowledge base, that collectively make up the prediction model usually known as decision algorithm. Rules can also be interpreted in various ways depending on the domain knowledge, data types(discrete or continuous) and in combinations. == RIPPER == Repeated incremental pruning to produce error reduction (RIPPER) is a propositional rule learner proposed by William W. Cohen as an optimized version of IREP.

Log-linear model

A log-linear model is a mathematical model that takes the form of a function whose logarithm equals a linear combination of the parameters of the model, which makes it possible to apply (possibly multivariate) linear regression. That is, it has the general form exp ⁡ ( c + ∑ i w i f i ( X ) ) {\displaystyle \exp \left(c+\sum _{i}w_{i}f_{i}(X)\right)} , in which the fi(X) are quantities that are functions of the variable X, in general a vector of values, while c and the wi stand for the model parameters. The term may specifically be used for: A log-linear plot or graph, which is a type of semi-log plot. Poisson regression for contingency tables, a type of generalized linear model. The specific applications of log-linear models are where the output quantity lies in the range 0 to ∞, for values of the independent variables X, or more immediately, the transformed quantities fi(X) in the range −∞ to +∞. This may be contrasted to logistic models, similar to the logistic function, for which the output quantity lies in the range 0 to 1. Thus the contexts where these models are useful or realistic often depends on the range of the values being modelled.

Environmental impact of AI

The environmental impact of the design, training, deployment and use of artificial intelligence includes the greenhouse gas emissions from generating electricity for data centres and computing hardware, operational and upstream water use, and material impacts from hardware manufacturing, mining and electronic waste. Estimating AI's environmental effects can be difficult because results depend on how impacts are measured, including whether accounting includes only model computation or also data-centre overhead, idle capacity, hardware manufacture, and local electricity supply. As these issues have received greater attention, governments and regulators have increasingly considered data-centre reporting requirements, energy-efficiency standards, and broader transparency measures for AI-related resource use. == Carbon footprint and energy use == AI-related energy use arises at multiple stages, including model training, fine-tuning, inference, storage, networking, and supporting infrastructure such as cooling and power conversion. === Individual level === Published estimates of energy use per AI request vary widely across models, tasks and measurement methods. A benchmark study presented at the 2024 ACM Conference on Fairness, Accountability, and Transparency found substantial differences between task types, with lower energy use for some text tasks and much higher energy use for image generation in the study's test conditions. In that benchmark, simple classification tasks consumed about 0.002–0.007 Wh per prompt on average (about 9% of a smartphone charge for 1,000 prompts), while text generation and text summarisation each used about 0.05 Wh per prompt; image generation averaged 2.91 Wh per prompt, and the least efficient image model in the study used 11.49 Wh per image (roughly equivalent to half a smartphone charge). First-party measurements in production environments have also been published. A 2025 Google study on Gemini assistant serving reported median per-prompt energy, emissions, and water-use estimates under the authors' accounting framework, while noting that different system boundaries can produce substantially different results. The study reported a median text-prompt estimate of about 0.24 Wh, which is roughly as much energy as watching nine seconds of television. The study also stated that software and infrastructure improvements reduced energy use by a factor of 33 and carbon emissions by a factor of 44 for a typical prompt over one year within the authors' framework. Researchers at the University of Michigan measured the energy consumption of various Meta Llama 3.1 models released in 2024 and found that smaller language models (8 billion parameters) use about 114 joules (0.03167 Wh) per response, while larger models (405 billion parameters) require up to 6,700 joules (1.861 Wh) per response. This corresponds to the energy needed to run a microwave oven for roughly one-tenth of a second and eight seconds, respectively. Comparisons between AI systems and human labour for specific tasks have produced mixed results and remain sensitive to assumptions about output quality, workload and system boundaries. A 2024 study in Scientific Reports reported 130 to 2900 times lower estimated carbon emissions for selected AI systems than for human writers and illustrators under its assumptions. A later Scientific Reports paper reported a counterexample for programming tasks under its assumptions, finding 5 to 19 times higher estimated emissions for the evaluated AI system than for human programmers on the benchmark used in that study. === System level === ==== Energy use and efficiency ==== AI electricity intensity depends not only on model architecture but also on hardware and facility efficiency. Data-centre operators commonly report Power usage effectiveness (PUE), which measures the ratio of total facility energy to IT equipment energy; a lower PUE indicates less overhead energy for cooling and other supporting infrastructure. Operators may also publish metrics and case studies on hardware efficiency, cooling systems and power sourcing. In its 2024 environmental report, Google stated that its 2023 total greenhouse gas emissions increased 13% year over year, primarily because of increased data-centre energy consumption and supply-chain emissions, while also reporting lower PUE than industry averages for its own facilities. The International Energy Agency has also reported that data centres remain a relatively small share of global electricity use overall, but that their local effects can be much more pronounced because demand is geographically concentrated. ==== Carbon footprint ==== At system level, AI contributes to rising electricity demand in data centres and related infrastructure. The International Energy Agency estimated that data centres used about 415 TWh of electricity in 2024, or around 1.5% of global electricity consumption, and projected that data-centre electricity use could rise to about 945 TWh by 2030, with AI identified as the main driver of that growth alongside other digital services. The carbon footprint of AI systems depends strongly on electricity sources, hardware efficiency, utilisation rates, and what stages are included in the accounting. Training large models can require substantial electricity, while total lifecycle impacts also depend on deployment scale and the amount of inference performed after training. Early analyses of frontier-model development reported rapid historical growth in training compute for selected systems, although later trends have depended on changes in model design, hardware and efficiency gains. Accounting methods that include upstream or embodied impacts, such as hardware manufacture and facilities construction, can materially affect estimates of AI-related emissions. === Decisions and strategies by individual companies === Large technology companies have reported that the expansion of AI and cloud infrastructure affects their sustainability targets, electricity demand, and resource use. Google, for example, attributed part of its emissions growth in 2023 to increased data-centre energy consumption and supply-chain emissions in its 2024 environmental report. Cloud and AI companies have also announced measures intended to reduce environmental impacts, including investment in more efficient hardware, low-carbon electricity procurement, alternative cooling systems, and water stewardship programmes. The extent, comparability, and third-party verification of such disclosures vary between firms and jurisdictions. == Water usage == Data centres can use water directly for cooling and indirectly through the water used in electricity generation, depending on the local energy mix. Public reporting on data-centre water use has often been inconsistent, making comparisons between operators and regions difficult. To standardise operational reporting, The Green Grid proposed the metric water usage effectiveness (WUE), defined as annual site water use divided by IT equipment energy use. WUE does not by itself measure local water stress, source sustainability, or all upstream water impacts. Studies of AI water use also distinguish between water withdrawal and water consumption. Research on AI-specific water use has argued that the water footprint of AI systems can be difficult to observe and may vary substantially by location, cooling design, and electricity source. A 2025 Communications of the ACM article summarised methods for estimating AI water footprints and emphasised the distinction between water withdrawal and water consumption. Li and colleagues estimated that global AI water withdrawal could reach 4.2–6.6 billion cubic metres in 2027 under the scenarios examined in their article. Using GPT-3, released by OpenAI in 2020, as an example, they estimated that training the model in Microsoft's U.S. data centres could consume about 700,000 litres of onsite water and about 5.4 million litres in total when offsite electricity-related water use was included; they also estimated that 10–50 medium-length GPT-3 responses could consume about 500 mL of water, depending on when and where the model was deployed. Published prompt-level estimates have also varied by system and accounting framework: the 2025 Google study on Gemini assistant serving reported a median text-prompt estimate of about 0.26 mL under its framework. Location can materially affect the significance of data-centre water use. Research on U.S. data centres found that one-fifth of servers' direct water footprint came from moderately to highly water-stressed watersheds, while nearly half of servers were fully or partially powered by plants located in water-stressed regions. A 2025 Reuters report, citing data from Verisk Maplecroft and NatureFinance, said that an average mid-sized data centre uses about 1.4 million litres of water per day for cooling and that Phoenix would experience a 32% increase in annual water stress if currently pl

Policy gradient method

Policy gradient methods are a class of reinforcement learning algorithms and a sub-class of policy optimization methods. Unlike value-based methods which learn a value function to derive a policy, policy optimization methods directly learn a policy function π {\displaystyle \pi } that selects actions without consulting a value function. For policy gradient to apply, the policy function π θ {\displaystyle \pi _{\theta }} is parameterized by a differentiable parameter θ {\displaystyle \theta } . == Overview == In policy-based RL, the actor is a parameterized policy function π θ {\displaystyle \pi _{\theta }} , where θ {\displaystyle \theta } are the parameters of the actor. The actor takes as argument the state of the environment s {\displaystyle s} and produces a probability distribution π θ ( ⋅ ∣ s ) {\displaystyle \pi _{\theta }(\cdot \mid s)} . If the action space is discrete, then ∑ a π θ ( a ∣ s ) = 1 {\displaystyle \sum _{a}\pi _{\theta }(a\mid s)=1} . If the action space is continuous, then ∫ a π θ ( a ∣ s ) d a = 1 {\displaystyle \int _{a}\pi _{\theta }(a\mid s)\mathrm {d} a=1} . The goal of policy optimization is to find some θ {\displaystyle \theta } that maximizes the expected episodic reward J ( θ ) {\displaystyle J(\theta )} : J ( θ ) = E π θ [ ∑ t = 0 T γ t R t | S 0 = s 0 ] {\displaystyle J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{t=0}^{T}\gamma ^{t}R_{t}{\Big |}S_{0}=s_{0}\right]} where γ {\displaystyle \gamma } is the discount factor, R t {\displaystyle R_{t}} is the reward at step t {\displaystyle t} , s 0 {\displaystyle s_{0}} is the starting state, and T {\displaystyle T} is the time-horizon (which can be infinite). The policy gradient is defined as ∇ θ J ( θ ) {\displaystyle \nabla _{\theta }J(\theta )} . Different policy gradient methods stochastically estimate the policy gradient in different ways. The goal of any policy gradient method is to iteratively maximize J ( θ ) {\displaystyle J(\theta )} by gradient ascent. Since the key part of any policy gradient method is the stochastic estimation of the policy gradient, they are also studied under the title of "Monte Carlo gradient estimation". == REINFORCE == === Policy gradient === The REINFORCE algorithm, introduced by Ronald J. Williams in 1992, was the first policy gradient method. It is based on the identity for the policy gradient ∇ θ J ( θ ) = E π θ [ ∑ t = 0 T ∇ θ ln ⁡ π θ ( A t ∣ S t ) ∑ t = 0 T ( γ t R t ) | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{t=0}^{T}\nabla _{\theta }\ln \pi _{\theta }(A_{t}\mid S_{t})\;\sum _{t=0}^{T}(\gamma ^{t}R_{t}){\Big |}S_{0}=s_{0}\right]} which can be improved via the "causality trick" ∇ θ J ( θ ) = E π θ [ ∑ t = 0 T ∇ θ ln ⁡ π θ ( A t ∣ S t ) ∑ τ = t T ( γ τ R τ ) | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{t=0}^{T}\nabla _{\theta }\ln \pi _{\theta }(A_{t}\mid S_{t})\sum _{\tau =t}^{T}(\gamma ^{\tau }R_{\tau }){\Big |}S_{0}=s_{0}\right]} Thus, we have an unbiased estimator of the policy gradient: ∇ θ J ( θ ) ≈ 1 N ∑ n = 1 N [ ∑ t = 0 T ∇ θ ln ⁡ π θ ( A t , n ∣ S t , n ) ∑ τ = t T ( γ τ − t R τ , n ) ] {\displaystyle \nabla _{\theta }J(\theta )\approx {\frac {1}{N}}\sum _{n=1}^{N}\left[\sum _{t=0}^{T}\nabla _{\theta }\ln \pi _{\theta }(A_{t,n}\mid S_{t,n})\sum _{\tau =t}^{T}(\gamma ^{\tau -t}R_{\tau ,n})\right]} where the index n {\displaystyle n} ranges over N {\displaystyle N} rollout trajectories using the policy π θ {\displaystyle \pi _{\theta }} . The score function ∇ θ ln ⁡ π θ ( A t ∣ S t ) {\displaystyle \nabla _{\theta }\ln \pi _{\theta }(A_{t}\mid S_{t})} can be interpreted as the direction in the parameter space that increases the probability of taking action A t {\displaystyle A_{t}} in state S t {\displaystyle S_{t}} . The policy gradient, then, is a weighted average of all possible directions to increase the probability of taking any action in any state, but weighted by reward signals, so that if taking a certain action in a certain state is associated with high reward, then that direction would be highly reinforced, and vice versa. === Algorithm === The REINFORCE algorithm is a loop: Rollout N {\displaystyle N} trajectories in the environment, using π θ t {\displaystyle \pi _{\theta _{t}}} as the policy function. Compute the policy gradient estimation: g i ← 1 N ∑ n = 1 N [ ∑ t = 0 T ∇ θ t ln ⁡ π θ ( A t , n ∣ S t , n ) ∑ τ = t T ( γ τ R τ , n ) ] {\displaystyle g_{i}\leftarrow {\frac {1}{N}}\sum _{n=1}^{N}\left[\sum _{t=0}^{T}\nabla _{\theta _{t}}\ln \pi _{\theta }(A_{t,n}\mid S_{t,n})\sum _{\tau =t}^{T}(\gamma ^{\tau }R_{\tau ,n})\right]} Update the policy by gradient ascent: θ i + 1 ← θ i + α i g i {\displaystyle \theta _{i+1}\leftarrow \theta _{i}+\alpha _{i}g_{i}} Here, α i {\displaystyle \alpha _{i}} is the learning rate at update step i {\displaystyle i} . == Variance reduction == REINFORCE is an on-policy algorithm, meaning that the trajectories used for the update must be sampled from the current policy π θ {\displaystyle \pi _{\theta }} . This can lead to high variance in the updates, as the returns R ( τ ) {\displaystyle R(\tau )} can vary significantly between trajectories. Many variants of REINFORCE have been introduced, under the title of variance reduction. === REINFORCE with baseline === A common way for reducing variance is the REINFORCE with baseline algorithm, based on the following identity: ∇ θ J ( θ ) = E π θ [ ∑ t = 0 T ∇ θ ln ⁡ π θ ( A t | S t ) ( ∑ τ = t T ( γ τ R τ ) − b ( S t ) ) | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{t=0}^{T}\nabla _{\theta }\ln \pi _{\theta }(A_{t}|S_{t})\left(\sum _{\tau =t}^{T}(\gamma ^{\tau }R_{\tau })-b(S_{t})\right){\Big |}S_{0}=s_{0}\right]} for any function b : States → R {\displaystyle b:{\text{States}}\to \mathbb {R} } . This can be proven by applying the previous lemma. The algorithm uses the modified gradient estimator g i ← 1 N ∑ n = 1 N [ ∑ t = 0 T ∇ θ t ln ⁡ π θ ( A t , n | S t , n ) ( ∑ τ = t T ( γ τ R τ , n ) − b i ( S t , n ) ) ] {\displaystyle g_{i}\leftarrow {\frac {1}{N}}\sum _{n=1}^{N}\left[\sum _{t=0}^{T}\nabla _{\theta _{t}}\ln \pi _{\theta }(A_{t,n}|S_{t,n})\left(\sum _{\tau =t}^{T}(\gamma ^{\tau }R_{\tau ,n})-b_{i}(S_{t,n})\right)\right]} and the original REINFORCE algorithm is the special case where b i ≡ 0 {\displaystyle b_{i}\equiv 0} . === Actor-critic methods === If b i {\textstyle b_{i}} is chosen well, such that b i ( S t ) ≈ ∑ τ = t T ( γ τ R τ ) = γ t V π θ i ( S t ) {\textstyle b_{i}(S_{t})\approx \sum _{\tau =t}^{T}(\gamma ^{\tau }R_{\tau })=\gamma ^{t}V^{\pi _{\theta _{i}}}(S_{t})} , this could significantly decrease variance in the gradient estimation. That is, the baseline should be as close to the value function V π θ i ( S t ) {\displaystyle V^{\pi _{\theta _{i}}}(S_{t})} as possible, approaching the ideal of: ∇ θ J ( θ ) = E π θ [ ∑ t = 0 T ∇ θ ln ⁡ π θ ( A t | S t ) ( ∑ τ = t T ( γ τ R τ ) − γ t V π θ ( S t ) ) | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{t=0}^{T}\nabla _{\theta }\ln \pi _{\theta }(A_{t}|S_{t})\left(\sum _{\tau =t}^{T}(\gamma ^{\tau }R_{\tau })-\gamma ^{t}V^{\pi _{\theta }}(S_{t})\right){\Big |}S_{0}=s_{0}\right]} Note that, as the policy π θ t {\displaystyle \pi _{\theta _{t}}} updates, the value function V π θ i ( S t ) {\displaystyle V^{\pi _{\theta _{i}}}(S_{t})} updates as well, so the baseline should also be updated. One common approach is to train a separate function that estimates the value function, and use that as the baseline. This is one of the actor-critic methods, where the policy function is the actor and the value function is the critic. The Q-function Q π {\displaystyle Q^{\pi }} can also be used as the critic, since ∇ θ J ( θ ) = E π θ [ ∑ 0 ≤ t ≤ T γ t ∇ θ ln ⁡ π θ ( A t | S t ) ⋅ Q π θ ( S t , A t ) | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=E_{\pi _{\theta }}\left[\sum _{0\leq t\leq T}\gamma ^{t}\nabla _{\theta }\ln \pi _{\theta }(A_{t}|S_{t})\cdot Q^{\pi _{\theta }}(S_{t},A_{t}){\Big |}S_{0}=s_{0}\right]} by a similar argument using the tower law. Subtracting the value function as a baseline, we find that the advantage function A π ( S , A ) = Q π ( S , A ) − V π ( S ) {\displaystyle A^{\pi }(S,A)=Q^{\pi }(S,A)-V^{\pi }(S)} can be used as the critic as well: ∇ θ J ( θ ) = E π θ [ ∑ 0 ≤ t ≤ T γ t ∇ θ ln ⁡ π θ ( A t | S t ) ⋅ A π θ ( S t , A t ) | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=E_{\pi _{\theta }}\left[\sum _{0\leq t\leq T}\gamma ^{t}\nabla _{\theta }\ln \pi _{\theta }(A_{t}|S_{t})\cdot A^{\pi _{\theta }}(S_{t},A_{t}){\Big |}S_{0}=s_{0}\right]} In summary, there are many unbiased estimators for ∇ θ J θ {\textstyle \nabla _{\theta }J_{\theta }} , all in the form of: ∇ θ J ( θ ) = E π θ [ ∑ 0 ≤ t ≤ T ∇ θ ln ⁡ π θ ( A t | S t ) ⋅ Ψ t | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=E_{\pi _{\theta }}\left[\su