Minutes

Resuming notes on the Cartif meeting of May 6 and 7/2014.
These notes report the strongholds of the discussion held in Cartif , possibly
updated in view of the subsequently exchanged e-mails.
First day: 4.00 p.m. 8.00 p.m.
1. Concrete methods and algorithms for fuzzy rules implementation (NTUA)
a. Cartif will translate knowledge found in D5.1 into something
machine readable (matlab, jfuzzy).
b. The fuzzy rules (as well as the reinforcement learning algorithm)
will be feed by the feedbacks to the previous recipe instantiations.
It is a dynamic process where feedbacks induce a new recipe on a
given task (or the most similar one) and the feedback to the latter
a new recipe more and so on along the exercise log of the
appliance. Of course, feedbacks are not the sole input to the
mentioned algorithms. The task specifications and user profile will
also be included in the input.
c. The output of the prototype will consist of a slow incremental
adaptation starting from current baseline recipes. A baseline
recipe is either the one supplied by the producer or the best one
implementing a task similar to the currently requested one, where
similarity metrics are to be tested. (At moment both recipes for
bread and wash have been uploaded on the Mongo DB residing in
the cloud server.)
d. Rules will be manually defined while the membership functions
will be learned (training). (A set of rules have been assessed by
UNIMI team and will be uploaded on cloud Mongo DB soon).
e. The training dataset will be constructed in the Cartif mockup
during the initial experimentation according to the agreed protocol
by Cartif, Unimi and (expectedly) Gorenje and will be used for the
ANFIS or any IFS training, on the partner convenience. All
partners are invited to formulate rules and training them. A daily
table will e published on the server reporting the results of the
most recent experiments.
f. The figure below resumes what has been agreed. The Input will be
compared to previous inputs and the most similar's corresponding
Recipe is returned. This recipe and the initial Input is fed to the IFS
which has been trained and rules are applied to the recipe. The
whole represents a run of the incremental learning cycle. Thus, on
the second run, the input is completed by the user feedback. (The
MongoBD infrastructure to manage these data and produce a
common base of inference for all partners is in progress.)
Hence the OHV server database must contain :
i. the reference to a file or a directory where all information
(fuzzy rules, membership functions, notes, matlab files) used
to move from one recipe to its evolution.
ii. on each new recipe the pointer to its ancestor.
2. Concrete methods and algorithms for reinforcement learning
implementation ( UPV)
It has been discussed the complementarity of the IFS methods w.r.t.
reinforcement learning. They have been surveyed the main kinds of
reinforcement learning algorithms. The Actor Critic family has been
agreed to be the most suitable in our case. Actor decisions are driven by
the fuzzy rules trained in the IFS framework plus wise random deviations.
Critics are directly collected from the eahouker feedbacks.
3. Concrete methods and algorithms for mining recipes from EDB
(UPV/CARTIF)
This point has been implicitly discussed in point 1. It has been outlined
that specific task parameters are not too huge. Namely for:
• Washing machine:
Textile, dirtiness degree, colour, weight (plus soap and softener
kinds)
• Bread maker
Flour, crust, weight (plus additional ingredients, such as nuts,
erbs,e tc)
To them, generic parameters attaining the user profile must be added.
The comparison methods through which to decree task similarities may
be drawn from the literature.
4. Selection of the basic recipes (CARTIF)
Cartif presents a fuzzy method to determine the recipe parameters as a
weighted sum of some baseline recipes, where weights come from the
membership degree of a given task to the corresponding baseline tasks. It
has been observed that a drawback of this procedure is the uniform
weighting of all parameters of a recipe, independently of their semantic. It
as been asked for an effort to reverse memberships specifically on the
parameters rather than on the entire recipe.
Second day: 8.30 a.m. 1.00 p.m.
5. drawing feedbacks from the mockup appliances (UNIMI)
UNIMI shows the washing machine and bread machine installed and
connected to the network in the Cartif Mockup.
6. appliance experiment protocols (CARTIF)
two kind of experiments have been decided with the following structure
of the input:
 for BM
o type of flour = [white, brown, sweet]
o crustiness = [3 level]
o amount = 600gr big 400 medium 250 small]
o At the training phase this quantization may prove suitable. At
recall phase we should be able to select any quantity of bread, in
turn belonging to the above sets through a membership function.
o Other parameters are marginal ingredients, such as salt and nuts.
Critical parameters such as water quantity and oil quantity are
fixed by the recipe
 for WM
o amount of clothes [5 kgr maximum 3kgr medium 1.5kgr minimum]
(Same consideration about the quantization)
o textile kind [cotton, wool, ...] (modeled as percentages of material
types (nylon, wool, etc.)), recalling some incompatibilities (say
woolens and silk)
o dirtiness degree [4 levels (3 standard plus stains)]
o colour [5 levels (according to what now happens with washing
programs): white, white + slight colored, colored, delicate, and full
colored]
o Additional parameters are the kind of soap and softener.
All the mentioned parameters have to be recorded in specific fields of the
DB.
A first phase of experimental campaign has been agreed consisting of one
month, two experiments per day per appliance. The operational workload
(charge/discharge the bread and washing machine) will be borne by
Cartif in this phase, while the involvement of Gorenje, the lead beneficiary
of WP7 has to be fixed for further phases and appliances. The plan of
experiments for this first phase is under Cartif responsibility, with the
collaboration of all partners (special contribution requested by Gorenje)
7. numerical experiment plans (NTUA,UPV)
For the first phase the following two goals have been confirmed
1.
Bread machine: recipes for 100 to 500 grams bread.
o Strategy: we point on one kind of bread (per time), using standard
mix of ingredients, apart for water. We fix water and cooking cycle
as a function of the desired bread quantity and standard reactions
of the user.
o Methodology: we use on the cloud EDB an ad hoc experiment table
(derived from the mentioned experiment plan) to manage the
experiments and the standard task request and eahouker feedback
collections. Two experiments per day. One people from Cartif and
one people from Milano will deal the experiments.
2. Washing machine: recipes for 5 kilos of clothes using the minimum
quantity of water and soap.
o Strategy: we point on one kind of payload, with a standard mix of
tissues. Same drying and soiling procedure after each wash. We
look for the minimum amount of water and powder which satisfy
the user.
o Methodology, same as for bread machine.