K.M. Wade

A survey of dairy calf management practices in Canada that affect animal welfare

There is growing interest among the public in farm animal welfare and a need for methods to assess animal welfare on farm. A survey on calf rearing practices that might affect dairy calf welfare was performed via a 1-h interview on 115 dairy farms (mean +/- SD: herd size=52.5+/-20.9 cows; milk production=8,697+/-1,153L) distributed throughout the province of Quebec. Despite frequent recommendations, many dairy producers continue to use management practices that increase the health risks of milk-fed calves. Major risk factors for poor calf welfare identified were 1) no use of calving pen in 51.3% of herds and low level of surveillance of calvings, especially at nighttime (once every 12h), 2) no disinfection of newborns navel in 36.8% of herds, and delayed identification and, hence, calf monitoring (3 d), 3) 15.6% of farms relied on the dam to provide colostrum and none checked colostrum quality or passive transfer of immunity, 4) dehorning and removal of extra teats proceeded at late ages (6.4 wk and 6.7 mo, respectively) and without adequate pain control, 5) use of traditional restrictive milk feeding and waste milk distributed to unweaned calves without precaution in 48.2% of herds, 6) abrupt weaning performed in 16.5% of herds, and 7) calves housed individually in 87.9% of herds, and most inappropriate housing systems (crate=27.0%, tie-stall=13.9%, attached against a wall=5.7%) remained. This risk factor assessment was the first step in an intervention strategy to improve calf welfare on dairy farms.

EFFECTS OF DATA PREPROCESSING ON THE PERFORMANCE OF ARTIFICIAL NEURAL NETWORKS FOR DAIRY YIELD PREDICTION AND COW CULLING CLASSIFICATION

The effect of data preprocessing on the learning ability of artificial neural networks was investigated with regard to the impact of distributing the input vectors uniformly with respect to the output categories in the training data set. The analyses were performed for neural networks dedicated to (1) dairy cow culling classification and (2) milk yield prediction. The two types of neural network used for culling classification were backpropagation and learning vector quantization. For yield prediction, backpropagation was used. The study was repeated with several architectures for both types of network. Preprocessing of data did not have a large impact on the general performance of the networks, but did affect the results for each output category. The effects were more pronounced in the categories containing less frequent events, for which the results always improved. For the categories with larger number of records, balancing the data degraded the results. The respective improvements and degradation of the results occurred for both prediction and classification, with the two types of neural networks, and with all architectures tested. However, the magnitude of the effects varied with the type of neural network and with the architecture. The results of this study indicate that, in general, the distribution of outputs influences the learning process of neural networks for both types of application. The results also suggest that the types of output distribution required for the training of neural nets may depend on the specifics of each problem.

NEURAL DETECTION OF MASTITIS FROM DAIRY HERD IMPROVEMENT RECORDS

A back-propagation artificial neural network was employed to detect clinical mastitis using a file of 460,474 test day records. Two data files were created to train the artificial neural networks, containing a relatively large (1:1) ratio and a relatively small (1:10) ratio in the incidence to non-incidence of clinical mastitis. These ratios were applied to each of two input file designs; one comprised variables that are traditional in the modeling of mastitis (e.g., age, stage of lactation and somatic cell count) and a second included additional variables (e.g., season of calving, milk components and conformation class). Results from analyses of relative operating characteristics indicated that artificial neural networks could discriminate between mastitic states with an overall accuracy of 86%. This discriminatory ability was subject to patterns that existed in the training data files but was not affected by differing proportions of mastitic records. However, differing proportions of mastitic records had some effect on the particular purpose of the artificial neural network being developed: training with a higher proportion of mastitic cases increased the ability of an artificial neural network in discriminating positive from negative cases. Similar effects were obtained by modifying the threshold value used to categorize the ANN output, which constitutes a much simpler approach than modifying the proportion of cases in training data sets. Additional variables had little effect on the prediction accuracy, but this lack of effect needs to be verified for optimal artificial neural network configuration, data preprocessing, and new sources of information.

Prediction of Cow Performance with a Connectionist Model

Since the main reason for disposal of dairy cows is low milk yield, implementation of an optimum selection program requires the prediction of cow performance with regard to production. The prediction of fat and protein content in milk are also rapidly becoming important factors for decisions related to breeding and herd policy. While, on average, traditional lactation models furnish good results, some improvement is possible when predicting the yield for an individual cow early in lactation. Artificial neural networks (ANNs), known to perform well in pattern recognition, may constitute an effective alternative to the traditional models. The objective of this research was to investigate how ANNs might be used to predict total milk, fat, and protein production for individual cows. Results indicated that ANNs generally performed at least as well overall as the model currently used by Canadian milk recording agencies, especially in the first third of lactation. This has important implications for early identification of superior animals. Predictions from both methods were relatively similar for the later stages of lactation. The addition of nontraditional data inputs such as average milk herd production and weight of cow improved the quality of prediction. Three different techniques were used to analyze the sensitivity of the ANN to different inputs, and their relative abilities are discussed. Results illustrate the potential effectiveness of ANNs in predicting milk yield and its composition and appear to justify further pursuit of this research.

Improving dairy yield predictions through combined record classifiers and specialized artificial neural networks

Abstract A multi-network system, comprising a neural classifier and two specialized neural predictors, was developed for prediction of milk yield from monthly records of Holstein dairy cattle. The classifier categorized records based on high- (≥9000 kg) and low-yields (

Design considerations for the implementation of multi-agent systems in the dairy industry

The objectives of this research were: (a) to perform a survey of current research in the area of multi-agent systems in order to learn more about how they could be designed and implemented; (b) to investigate the feasibility of such an approach for agriculture, based on an integration of currently existing technologies; and, more specifically (c) to assess the potential of a multi-agent approach in the context of decision support for dairy production. The results of this work highlighted a number of key concepts in multi-agent system design, including the importance of selecting an appropriate system architecture for agent coordination (e.g., peer-to-peer, federated, or blackboard-based) and the need for well-defined agent communication methods (language and ontology). Alternative technologies, used in the implementation of multi-agent systems (e.g., communication protocols and distributed computing methods), were also explored. Lastly, a case study was carried out, in which some of the discussed technologies were tested and implemented to create a multi-agent heifer management system. The system consists of two different types of agents and several databases, and was implemented on a PC-based network. The agents work together to synthesize data about heifer development from different sources and to present this to the user in a graphical format. The system demonstrates the feasibility of applying an agent-based approach, using currently available technology, to problems such as dairy herd management in which a distributed decision-support solution is often required. It is concluded that the constraints for the implementation of multi-agent systems do not appear to be of a technological nature; the challenge seems to be more one of defining and accepting a common ontology and communication language by members of a given industry. In addition, large-scale distributed systems will require sophisticated agent-coordination methods to ensure robust and efficient operation.

EFFECTS OF LEARNING PARAMETERS AND DATA PRESENTATION ON THE PERFORMANCE OF BACKPROPAGATION NETWORKS FOR MILK YIELD PREDICTION

The effect of four different learning parameters as well as the method of data presentation was investigated with regard to the performance of backpropagation neural networks in predicting milk yield. The parameters examined were: (1) the learning rates of the hidden and output layer, (2) momentum, (3) epoch size, and (4) a binary/bipolar data presentation. The modified values of each parameter included the extremes and commonly used levels found in the literature. Modifications were made one at a time, and network training was repeated with different sets of initial weights. In addition, combinations of bipolar data presentation with different epoch sizes were also formed to study their combined effect. Although all the networks learned their training data quite well, and quite similarly, some differences were observed in the results of test data. The most notable effects were detected when the epoch assumed its extreme sizes, and when a bipolar data representation was used. The lowest network root mean square errors corresponded equally to the smallest epoch size, a bipolar data presentation, and their combination. Increasing the value of the learning rate in the hidden layer tended to improve network performance. Conversely, large learning rates in the output layer tended to increase the network error. With regard to different values of the momentum, the results were quite similar. The results of repetitions, for all parameters, revealed that the initial weights influenced network performance. Results obtained in this study suggest that the users of artificial neural networks should pay attention to the values of different learning parameters and the method of data presentation. They should also carry out several repetitions, using different initial weight values, in order to optimize the network results.

FUZZY SET-BASED ANALYTICAL TOOLS FOR DAIRY HERD IMPROVEMENT

Dairy producers receive large amounts of data from Dairy Herd Improvement agencies on a monthly basis, not only for the purpose of supplying summary statistics, but also as an aid in the detection of management weaknesses. The latter task can be carried out more efficiently by producers and their advisors if aided by decision-support systems, the role of which is to do preliminary data processing. Two software programs were developed to carry out such pre-processing for individual cows in terms of milk yield and persistency. Specifically, these programs analyze both traits as a function of parity, stage of lactation, and herd average production level, and are based on fuzzy sets, which furnish qualitative assessments of the deviations from standard values. This approach has proved to work well, and may be extended to other factors such as peak yield, body condition score and somatic cell count.

Induction and evaluation of decision trees for lactation curve analysis

Abstract Machine learning has been identified as a promising approach to knowledge-based system development. This study focused on the use of decision-tree induction for knowledge acquisition to filter individual-cow lactations for group-average lactation curve analysis. Data consisted of 1428 cases, classified by a dairy-nutrition specialist as outliers (34 cases) or non-outliers. The classification performance was estimated through 10-fold cross validation. A relative operating characteristic curve was used to visualize the achievable range of trade-offs between correctly classifying positive and negative cases. A series of three final decision trees with increasing tendency of classifying a lactation as outlier was induced from the entire data set. For these trees, the expected true positive rates were 52, 68 and 92%, at false positive rates of 1.5, 3.5 and 8.6%, respectively. However, due to the low prevalence of outlier lactations (cases), this performance was associated with many false positives. The performance of individual decision nodes was tested against the entire data set to identify potentially counter-intuitive nodes resulting from overspecialization to the training data. The specialist reviewed the final trees and adjusted two decision nodes. This study suggests that although the input from a domain specialist remains important, decision-tree induction is a useful technique to support knowledge acquisition involved in the removal of outlier lactations.

Performance analysis for machine-learning experiments using small data sets

Abstract Machine-learning techniques are increasingly used to deal with a variety of problems in agriculture. However, challenges with the application of machine-learning, such as analyzing the performance achieved through learning from small data sets, still remain. This study focused on using graphical and statistical techniques to analyze the results of machine-learning experiments involving data preprocessing and algorithm tuning. Data consisted of 1428 cases that were classified by a dairy-nutrition specialist as outliers (34 cases) or non-outliers. The performance of classifiers, generated with decision-tree induction, was estimated through ten-fold cross validation. Relative operating characteristic (ROC) curves were used to visualize the achieved trade-offs between correctly classifying positive and negative cases. A performance index, representing the mean true positive rate of these curves for a limited range of false positive rate values, was developed to facilitate comparison among classification schemes. Analysis of variance (ANOVA) was used to determine whether real differences existed for the expected performance on new data among the different combinations of data preprocessing and algorithm configurations evaluated in this study. In terms of data preprocessing, randomly assigning herds to the folds of the cross validation did not perform significantly differently from assigning cases to folds, while using a special value to indicate irrelevant attribute values significantly improved the performance over treating these values as unknown. Tuning the configuration of the decision-tree induction algorithm significantly improved the classification performance. The application of ten-fold cross validation in combination with ROC curves and ANOVA was found to be useful in analyzing the results of machine-learning experiments involving decision-tree induction and small data sets. These methods could also be used with other machine-learning techniques such as artificial neural networks and instance-based learning.