Материал: искусственный интеллект

84 A. Wichert and C. Moreira

We can solve the Equation for the phase θ¬ or θ resulting in two possible cases. For θ¬ we get

If the preceding constraint is not valid, then we solve the equation for the

For equality in the constraint p(y) = p(¬y) both cases become equal.

Probability Waves Are Smaller Equal One. We assume without loss of

86 A. Wichert and C. Moreira

Fig. 3. Probability waves for the experiments described in Tables 1 and 2. In plots (a) - (d) the waves p(¬y, θ¬ ) are around p(¬y), (for (e) see Fig. 2). In the plots (i) - (iv) the waves p(y, θ) are around p(y). Additionally the values psub (¬y) and psub (y) are indicated by a line. Note that the curves in the plots (i) – (iv) overlap.

3 Law of Maximal Uncertainty

How can we choose the phase θ for a probability value that reflects the case of not knowing what the correct value is and without losing information about the probability waves? We answer this question by proposing the law of maximal uncertainty is based on two principles, the principle of entropy and the mirror principle.

3.1Principle of Entropy

For the case in which both interval do not overlap

the values of the waves that are closest to the equal distribution are chosen. By doing so the uncertainty is maximised and the information about the probability wave is not lost. The principle of maximum entropy states that the probability distribution which best represents the current state of knowledge is the one with largest entropy [11–13]. In the case of a binary event the highest entropy corresponds to an equal distribution

H = −p(y) · log2 p(y) − p(¬y) · log2 p(¬y) = −log2 · 0.5 = 1 bit. (39)

For p(y) ≤ p(¬y) the closest values to the equal distribution are for θ = 0, the ends of the interval for

an equal distribution maximises the uncertainty but loses the information about the probability wave. To avoid the loss we do not change the entropy of the system we use the positive interference as defined by the law of balance. When the intervals overlap the positive interference is approximately of the size of smaller probability value since the arithmetic and geometric means approach each other, see Eq. 26. We increase uncertainty by mirror the “probability values”. For the case p(y) ≤ p(¬y) we assume

Table 3 summarises the intervals that describe the probability waves, the resulting probabilities pq that are based on the law of maximal uncertainty, the subjective probability and the classical probability values. We compare the results that

Table 3. Probability waves, the resulting probabilities pq that are based on the law of maximal uncertainty, the subjective probability and the classical probability values. Entries (a)–(e) are based on the principle of entropy and entries (i)–(iv) are based mirror principle.

are based on probability waves and the law of maximal uncertainty with previous works that deal with predictive quantum-like models for decision making, see Table 4. The dynamic heuristic [18] is used quantum-like bayesian networks. Its parameters are determined by examples from a domain. On the other hand In the Quantum Prospect Decision Theory the values need not to be adapted to a domain. The quantum interference term is determined by the Interference Quarter Law. The quantum interference term of total probability is simply fixed to a value equal to 0.25 [22].