The test process for semiconductor devices involves generation and application of test patterns, failure logging and diagnosis. Traditionally, most of these activities cater for all possible faults without making any assumptions about the actual defects present in the circuit. As the size of the circuits continues to increase (following the Moore's Law) the size of the test sets is also increasing exponentially. It follows that the cost of testing has already surpassed that of design and fabrication. The central idea of our work in this dissertation is that we can have substantial savings in the test cost if we bring the actual hardware under test inside the test process's various loops -- in particular: failure logging, diagnostic pattern generation and diagnosis. Our first work, which we describe in Chapter 3, applies this idea to failure logging. We modify the existing failure logging process that logs only the first few failure observations to an intelligent one that logs failures on the basis of their usefulness for diagnosis. To enable the intelligent logging, we propose some lightweight metrics that can be computed in real-time to grade the diagnosibility of the observed failures. On the basis of this grading, we select the failures to be logged dynamically according to the actual defects in the circuit under test. This means that the failures may be logged in a different manner for devices having different defects. This is in contrast with the existing method that has the same logging scheme for all failing devices. With the failing devices in the loop, we are able to optimize the failure log in accordance with every particular failing device thereby improving the quality of diagnosis subsequently. In Chapter 4, we investigate the most lightweight of these metrics for failure log optimization for the diagnosis of multiple simultaneous faults and provide the results of our experiments. Often, in spite of exploiting the entire potential of a test set, we might not be able to meet our diagnosis goals. This is because the manufacturing tests are generated to meet the fault coverage goals using as fewer tests as possible. In other words, they are optimized for `detection count' and `test time' and not for `diagnosis'. In our second work, we leverage realtime measures of diagnosibility, similar to the ones that were used for failure log optimization, to generate additional diagnostic patterns. These additional patterns help diagnose the existing failures beyond the power of existing tests. Again, since the failing device is inside the test generation loop, we obtain highly specific tests for each failing device that are optimized for its diagnosis. Using our proposed framework, we are able to diagnose devices better and faster than the state of the art industrial tools. Chapter 5 provides a detailed description of this method. Our third work extends the hardware-in-the-loop framework to the diagnosis of scan chains. In this method, we define a different metric that is applicable to scan chain diagnosis. Again, this method provides additional tests that are specific to the diagnosis of the particular scan chain defects in individual devices. We achieve two further advantages in this approach as compared to the online diagnostic pattern generator for logic diagnosis. Firstly, we do not need a known good device for generating or knowing the good response and secondly, besides the generation of additional tests, we also perform the final diagnosis online i.e. on the tester during test application. We explain this in detail in Chapter 6. In our research, we observe that feedback from a device is very useful for enhancing the quality of root-cause investigations of the failures in its logic and test circuitry i.e. the scan chains. This leads to the question whether some primitive signals from the devices can be indicative of the fault coverage of the applied tests. In other words, can we estimate the fault coverage without the costly activities of fault modeling and simulation? By conducting further research into this problem, we found that the entropy measurements at the circuit outputs do indeed have a high correlation with the fault coverage and can also be used to estimate it with a good accuracy. We find that these predictions are accurate not only for random tests but also for the high coverage ATPG generated tests. We present the details of our fourth contribution in Chapter 7. This work is of significant importance because it suggests that high coverage tests can be learned by continuously applying random test patterns to the hardware and using the measured entropy as a reward function. We believe that this lays down a foundation for further research into gate-level sequential test generation, which is currently intractable for industrial scale circuits with the existing techniques. / Ph. D.