Brain Computer Interfaces (BCI) are devices that translate acquired neural signals to command and control signals. Applications of BCI include neural rehabilitation and neural prosthesis (thought controlled wheelchair, thought controlled speller etc.) to aid patients with disabilities and to augment human computer interaction. A successful practical BCI requires a faithful acquisition modality to record high quality neural signals; a signal processing system to construct appropriate features from these signals; and an algorithm to translate these features to appropriate outputs. Intracortical recordings like local field potentials provide reliable high SNR signals over long periods and suit BCI applications well. However, the non-stationarity of neural signals poses a challenge in robust decoding of subject behavior. Most BCI research focuses either on developing daily re-calibrated decoders that require exhaustive training sessions; or on providing cross-validation results. Such results ignore the variation of signal characteristics over different sessions and provide an optimistic estimate of BCI performance. Specifically, traditional BCI algorithms fail to perform at the same level on chronological data recordings. Neural signals are susceptible to variations in signal characteristics due to changes in subject behavior and learning, and variability in electrode characteristics due to tissue interactions. While training day-specific BCI overcomes signal variability, BCI re-training causes user frustration and exhaustion. This dissertation presents contributions to solve these challenges in BCI research. Specifically, we developed decoders trained on a single recording session and applied them on subsequently recorded sessions. This strategy evaluates BCI in a practical scenario with a potential to alleviate BCI user frustration without compromising performance. The initial part of the dissertation investigates extracting features that remain robust to changes in neural signal over several days of recordings. It presents a qualitative feature extraction technique based on ranking the instantaneous power of multichannel data. These qualitative features remain robust to outliers and changes in the baseline of neural recordings, while extracting discriminative information. These features form the foundation in developing robust decoders. Next, this dissertation presents a novel algorithm based on the hypothesis that multiple neural spatial patterns describe the variation in behavior. The presented algorithm outperforms the traditional methods in decoding over chronological recordings. Adapting such a decoder over multiple recording sessions (over 6 weeks) provided > 90% accuracy in decoding eight movement directions. In comparison, performance of traditional algorithms like Common Spatial Patterns deteriorates to 16% over the same time. Over time, adaptation reinforces some spatial patterns while diminishing others. Characterizing these spatial patterns reduces model complexity without user input, while retaining the same accuracy levels. Lastly, this dissertation provides an algorithm that overcomes the variation in recording quality. Chronic electrode implantation causes changes in signal-to-noise ratio (SNR) of neural signals. Thus, some signals and their corresponding features available during training become unavailable during testing and vice-versa. The proposed algorithm uses prior knowledge on spatial pattern evolution to estimate unknown neural features. This algorithm overcomes SNR variations and provides up to 93% decoding of eight movement directions over 6 weeks. Since model training requires only one session, this strategy reduces user frustration. In a practical closed-loop BCI, the user learns to produce stable spatial patterns, which improves performance of the proposed algorithms. / text