Lecture 1.1.1: Monotone Disjunctions

The learning process involves starting with an initial monotone disjunction, represented mathematically as:

[ \text{Initial State} = x_1 \lor x_2 \lor \cdots \lor x_n ]

This particular disjunction is notable because it evaluates to true for any string input except the 'all zeros' string.

Processing a Challenge:When the learner receives a challenge, such as the bitstream 0110010, it assesses the bit positions. For example, since there are '1's at the second and third positions, the learner concludes, based on its initial state, that the output for this input should be '1'.

However, if the teacher indicates that this response was incorrect, the learner deduces that any literals present in its response (like x2 and x3) cannot be part of the true monotone disjunction. The learner evaluates this based on the principle that if x2 were indeed part of the disjunction, the output provided would have been correct; since it was not, x2 is eliminated:

• Elimination Process:

  • By responding '1' when the correct answer is '0', the learner gains information that x2 cannot be in the disjunction.

  • It also learns that because it assumed the presence of x3 (which was also '1' in the input), x3 must also be eliminated.

  • This logical deduction continues for any additional literals implicated in the mistake—if x6 is also validated as incorrect, it too is removed from consideration.

Implication of Mistakes:For every wrong response, at least one literal is removed from the initial disjunction. This iterative elimination process ensures that with each mistake made, the state of the learner’s understanding improves. The maximum number of mistakes the learner can make is bounded by the number of literals (n) in the initial state. Thus, the learning process becomes more refined with repeated challenges, as the learning algorithm systematically narrows down the possible correct disjunctions.

Challenge Complexity:Even though the number of potential challenges grows exponentially (2^n possible combinations for n literals), the efficiency of the learner’s algorithm is confirmed by its maximal mistake count, which is linear in relation to the number of literals. This efficiency highlights the efficacy of the monotone disjunction-based learning in dealing with complex binary input scenarios.