These papers covered nine separate sports. Though Naive Bayes and Bayesian Networks share common foundations, a decision was made to keep them separate since the latter includes the ability to incorporate domain knowledge to a larger degree, which was a motivating factor for its usage in several studies. What are some of the defining characteristics that can be drawn from successful studies? Then, in section 3, we review the literature in machine learning for sport result prediction and also present tabular summaries of studies by sport. Methods which combined several different families of machine learning algorithm into a single decision making architecture were termed Ensemble. Machine Learning algorithms were normalized by grouping them into families of algorithms in order to identify trends in their usage patterns. Purucker (1996) used an ANN as well as unsupervised learning techniques to predict the results of US National Football League (NFL) football matches in 1994, using data from weeks 11 to 16 of the competition (90 matches). The dataset was derived from three football teams and two Rugby teams, across three seasons. Reed Cincinnati Bengals Gameday & Tailgating O’Donoghue (2005) built seven predictive models (including models based on Multiple Linear Regression, an ANN and Discriminant Analysis) using seven predictor variables, to predict results in English Premier League (EPL) Soccer and Premiership Rugby in England.
This atlas facilitates group-wise statistical analysis for the assessment. This includes game outcome prediction, measurement and evaluation of player performance, analysis of rules and adjudication, and within-game strategy. The expert BN was found to provide the best performance, achieving 59.2% accuracy in predicting a home win, away win or a draw (a 3-class problem). Expert knowledge was incorporated while constructing a Bayesian Network (BN) model, and it was found that including such knowledge can result in strong performance, especially when the sample size is small (which is often the case with sports). A total of 11 inputs were used in the ANN for each game, and the Levenberg Marquadt (LM) routine was used to train the network. The SOM provided the best performance among the unsupervised methods, but it generally could not match the performance of the ANN. The Multi-class Classifier had the best prediction accuracy of 55%. For future work the author considered including more features such as: yellow/red cards, the number of players each team has, their managers, their player budget and their home ground capacity. An ANN with BP was used, and the features included in the model were: total yardage differential, rushing yardage differential, time in possession differential, turnover differential, a home or away indicator, home team outcome and away team outcome.
The most successful computer-based model was the Simulation Model. The computer-based models predicted between 39 and 44.5 of the 48 matches correctly, while the 42 human experts correctly predicted an average of 40.7 matches. Furthermore, it was mentioned that aggregations or ensembles of classifiers (e.g., voting) could be investigated, and that automatic feature selection methods should be used rather than manual human selection. Unsupervised learning methods were also applied, in particular: the Hamming, Adaptive Resonance Theory (ART), and Self Organizing Map (SOM) methods. Our focus is on the application of ML — fuzzy methods. 2020), but is not the focus of this review. 2020) and Rudrapal et al. Although we describe this model in terms of two teams accumulating points, it can in principle be generalized to other forms of competition. Inverse reinforcement learning (IRL) (Ng & Russell, 2000; Abbeel & Ng, 2004) would infer reward functions that promote the observed behavior in demonstrations, which can then be used in model-free RL. The statistics of the gameplay for the two agents playing together against the traditional game AI agents are shown in Table 6. While the second agent is trained with the same reward function as the first one, it is trained in a different environment as the teammate is now the offensive DQN agent trained in the previous experiment rather than the traditional game AI agent.
Finally, as shown in the lowest part of Table II, the results of the proposed method outperform the results of previous work, as well as those of the clustering-based baseline. For each tile the priority can be assigned in multiple ways, which gives our method the flexibility to adapt to different priority models. As such, we can apply a coarse-to-fine approach where we first align the data to a coarse template, and then using this initial alignment we can partition the data into finer states which provide templates which allow us to find a better alignment. The authors then cycled back to the data preparation phase of the CRISP-DM framework to see if adding additional variables could improve results. The authors applied the Cross-Industry Standard Process for Data Mining (CRISP-DM) framework (Wirth & Hipp, 2000) as their experimental approach, and investigated how the successful shot percentage in six different regions of the court affected match results.