We report here some improvements to be included in the future versions of
the bp-ct code.

We will introduce a probe to each agent, to allow the direct inspections
of agent variables while the model is running. An important improvement
will be that of the automation of the production of the CT formulas mainly
on the basis of definitions of actions and effects (at present the formulas
containing the CT rules are handwritten in the Interface code). Finally we
will operate to allow the automatic or quasi-automatic substitution of
neural networks with classifier systems or with more general genetic
algorithms.

We think that it would be useful to develop the General Hypothesis about
agent coherence (see Section 'The Cross-Target (CT) Method' of the paper
quoted in the HowTo file) also in other ways, to verify the reproducibility
of our results in other contexts. We have to introduce algorithms capable
of the same performances in order to reproduce short and long term learning, 
without the aid of ANN. Our algorithm must be able to modify its outputs in
a smooth way, following cross-target suggestions about actions and guesses
of action effects, to self-develop the behavioral skills of the acting and
adapting phase. The algorithm - may be not the same - has also to develop a
strong mapping ability between input and output (target) vectors, to produce
the same behavioral results without further learning.
