To typically off street (and digging traced back based 7307 to the problem to predicting ordinary written language. It requires either case vast, real-world knowledge. estimated that the information content of written case-insensitive English without punctuation is 0 to 1 bits per character, based on experiments which human subjects guessed successive characters text with the help of letter n-gram frequency tables and dictionaries. The uncertainty is due not much to variation subject matter and human skill as it is due to the fact that different probability assignments lead to the same observed guessing sequences. Nevertheless, the best text compressors are only now compressing near the upper end of this range. Legg and Hutter proposed the second definition, universal intelligence, to be far more general than Turing's human intelligence. They consider the problem of reward-seeking agents completely arbitrary environments described by random programs. this model, agent communicates with environment by sending and receiving symbols. The environment also sends a reinforcement or reward signal to the agent. The goal of the agent is to maximize accumulated reward. intelligence is defined as the expected reward over all possible environments, where the probability of each environment described by a program M is algorithmic, proportional to 2 -|M|. Hutter proved that the optimal strategy for the agent is to guess after each input that the distribution over M is dominated by the shortest program consistent with past observation. Hutter calls this strategy AIXI. It is, of course, is just our uncomputable compression problem applied to a transcript of past interaction. AIXI also be considered a formal statement and proof of Occam's Razor. The best predictor of the future is the simplest or shortest theory that explains the past. There is no such thing as universal compression, recursive compression, or compression of random data. Most strings are random. Most meaningful strings are not. Compression modeling coding. Coding is a solved problem. Modeling is provably not solvable. Compression is both and artificial intelligence problem. The key to compression is to understand the data you want to compress. A data compression benchmark measures compression ratio over a data set, and sometimes memory usage and speed on a particular computer. Some benchmarks evaluate size only, order to avoid hardware dependencies. Compression ratio is often measured by the size of the compressed output file, or bits per character meaning compressed bits per uncompressed byte. either case, smaller numbers are better. 8 bpc means no compression. 6 bpc means 25% compression or 75% of original size. Generally there is a 3 way trade off between size, speed, and memory usage. The top ranked compressors by size require a lot of computing resources. The Calgary corpus is the oldest compression benchmark still use. It was created 1987 and described a survey of text compression models 1989 It consists of 14 files with a total size of 3,622 bytes as follows: 111 BIB ASCII text refer format 725 bibliographic references. 768 BOOK1 unformatted ASCII text Hardy: Far from the Madding Crowd. 610 BOOK2 ASCII text troff format Witten: Principles of Computer Speech. 102 GEO 32 bit numbers floating point format seismic data. 377 NEWS ASCII text USENET batch file on a variety of topics. 21 OBJ1 executable program compilation of PROGP. 246 OBJ2 Macintosh executable program Knowledge Support System. 53 PAPER1 troff format Witten, Neal, Cleary: Arithmetic Coding for Data Compression. 82 PAPER2 troff format Witten: Computer security. 513 PIC 1728 x 2376 bitmap image text French and line diagrams. 39 PROGC Source code C compress v4. 71 PROGL Source code Lisp system software. 49 PROGP Source code Pascal program to evaluate PPM compression. 93 TRANS ASCII and control characters transcript of a terminal session. The struture of the corpus is shown the diagram below. Each pixel represents a match between consecutive occurrences of a string. The color of the pixel represents the length of the match: black for 1 byte, red for 2, green for 4 and blue for 8. The horizontal axis represents the position of the second occurrence of the string. The vertical axis represents the distance back to the match on a logarithmic scale. String match structure of the Calgary corpus The horizontal dark line at around 60 BOOK1 is the result of linefeed characters repeating at this regular interval. Dark lines at 1280 and 5576 GEO are due to repeated data block headers. The dark bands at multiples of 4 are due to the 4 byte structure of the data. PIC has a dark band at 1 due to runs of zero bytes, and lines at multiples of 216, which is the width of the image bytes. The blue curves at the top of the image show matches between different text files, namely BIB, BOOK1, BOOK2, NEWS, PAPER1, PAPER2, PROGC, PROGL, PROGP and TRANS, all of which contain English text. Thus, compressing these files together can result better compression because they contain mutual information. No such matches are seen between GEO, OBJ2, and PIC with other files. Thus, these files should be compressed separately. Other structure can be seen. For example, there are dark bands OBJ2