Hey DedeHai,

could you use 4 Pins of the ATMega for an multiplexing system?
If you use 1 Pin as an "Slave Select" and every slave get the same 
startup adress, u can count up the multiplexer and activate the 
i2c-interface from the slave by multiplexer-state. After that, you send 
a new adress over the bus, and the selected slave send an acknowledge. 
If the master dont get that, the selected slot is unused.

Thanks for your answers.

@Sepp:
It certainly would be an elegant way, but the bus is represented by 
cables between the devices. It would mean that instead of using a 3 
conductor cable (and plug) I would have to use 4 more conductors for the 
"select" signals.

@hans
As far as I understand the I2C protocol, a device can only pull the bus 
lines to low, so a 0 always overwrites a 1 if two masters transmit at 
the same time. so the one sending more zeroes will win the battle. The 
hardware of the attiny24 has a data collision flag which is set when the 
transmitted bit and the actual pin status are not the same. Like this it 
should be possible to do what I want. But you are right about the 
problem with two identical addresses.

Here are a few additional thoughts to that:

-another random number is generated (by reading a floating analog 
input), representing a delay cycle. Each attiny waits until the delay is 
up, then tries to transmit its address to the master. if a "start 
condition" (meaning another device is using the bus) occurred during the 
wait cycle (generates an interrupt) the device waits until a stop 
condition occurs, then continues with the wait cycle. like this, the 
risk of undetected data collisions can be minimized and two devices with 
the same address can potentially be detected. however it is still 
possible that two devices get the same addreass and the same delay time, 
resulting again in an undetected slave.

-once the atmega has detected the slaves, their addresses can be stored 
in eeprom, at the next startup a check byte can be sent out to all 
slaves, if a slave does not get the check byte it will transmitt it's 
address after this status check.

-To decrease the risk of having two slaves with the same address, 10 
random bytes could be sent by each attiny, resulting in 80 random bits 
that would have to be exactly the same on two (or more) attinys or a 
collision could be detected, so the attiny then can detect that another 
device (with the same address) is using the bus and then change its 
address again. The more random bytes transmitted, the lower the risk of 
having multiple devices with the same address. The downside with this is 
that it uses up more time and you never can be 100% sure.


Any good ideas on how to make 100% sure two slaves do not have the same 
address?

Why do you need 7? you only need 4 Wires for I2C, ChipSelect and GND.
Just the ATMEGA need 4 extra wires for controll the multiplexer IC, for 
example the 74154N. Then the slaves just have to check this one wire to 
activate there I2C inteface with the startup adress, and get a new 
adress from the mega. After that, the mega count up the adress and the 
multiplexer and write the next one to the tinys. No random adresses, and 
you can store the slave adresses in the flash of the mega.

In the other case, deactivate the i2c interface and use the lines in I/O 
state. Every slave try to pull down the data line first(after your 
random time). To check up that the winning slave is the only one, they 
read the clock line and pull it down after that. Then the mega goes in 
master state and the slave use an standard adress. The master (mega) 
send an new adress. After a while, the other slave wake up and retry.
But, for me, its realy undefined and tricky to use such a solution.

Also quite simple:
Programm an software SPI, and switch after sending the adresses to I2C.